BUILD DOCS

PROMPT PACKS: Prayer Delivery Offer Engine

prayer/PROMPT-PACKS.md

PROMPT PACKS: Prayer Delivery Offer Engine

Production asset for the fulfilment pipeline. Internal document. Brand v1: St. Peter's Basilica, Vatican (st_peters_confessio). Wave 1: Western Wall, Jerusalem (western_wall) and Fatima (fatima).

Every deliverable is produced by a two-pass, donor-conditioned run: Pass 1 regenerates the site from a real donor photograph, Pass 2 puts a hand holding the buyer's printed prayer card into the foreground. Card-in-hand is the default composition for every brand. The card itself is rendered, not generated — CARD-SPEC.md.

Image models only: GPT Image class and Nano Banana Pro class. Nothing else in the path.


1. PIPELINE OVERVIEW

1.1 The two passes

Stage Input Output Why
Card, printed first name + prayer text the finished card PNG, 1240x1748 at 300 dpi (105 x 148 mm) the card is a print, identical on every order and printable as a mail piece
Pass 1, scene donor reference image + scene prompt clean site photo, no card, no people, no hands the donor is what makes the site look like itself
Pass 2, card in hand Pass-1 output + the rendered card + a hand reference final deliverable: a hand holding the printed card in the foreground, site behind the card is the product

The card comes first, and it is deterministic. @prayer-engine/card typesets the buyer's own name and prayer onto our card and encodes a PNG — no model call, no drift between renders, and at 300 dpi so the identical file can go to a print house. Pass 2's relationship to it is to hold it up and photograph it. Pass 2 writes no text at all, so there is no name slot, no prayer slot and no spelling to get wrong.

Pass 1 conditions on the location's DONOR reference image, a real photograph of the site sourced from Pinterest, stored at brands/<slug>/assets/donor/<location_key>.jpg and declared in BrandConfig.location.donorRef. The prompt tells the model to match the reference's composition, materials and light, then re-render it under the order's randomized conditions. The donor is never shipped and never appears in the output; it is the geometry and material anchor.

Pass 2 is an edit on the Pass-1 frame. It adds one hand entering the lower frame holding the buyer's printed card — attached as a second image, complete — and changes nothing else. The buyer's prayer is capped at 320 characters by BrandConfig.prayer.maxChars, enforced at the survey, again at the API and again at the renderer, and the renderer's fit ladder sets the whole cap at a legible size. There is no long-prayer path and no excerpting step.

1.2 Randomization axes (per order)

Every order gets a different moment. Six axes, all derived from the order's variant_seed so a run is reproducible and a re-run of the same job gives the same photo.

Axis Values Derivation
Time of day morning, afternoon, evening, night tod = variant_seed mod 4
Light soft diffuse, raking directional, pooled flame light = (variant_seed >> 2) mod 3
Camera angle the location's scene variants (Section 6) angle = (variant_seed >> 4) mod n_variants
Hand H1 / H2 / H3 (skin tone, age, sleeve) hand = (variant_seed >> 7) mod 3
Card variant E1 / E2 / E3 card = (variant_seed >> 9) mod 3
Hand side left / right side = (variant_seed >> 11) mod 2

Every axis reads a different bit range on purpose. Side takes bits from >> 11 and not the low bit, because the low bit already drives time of day: read the same bit twice and morning and evening always enter from the left while afternoon and night always enter from the right, which throws away half the intended variety.

variant_seed is a 32-bit signed integer (prayers.variant_seed integer, OPS §3) stored on the prayer row at enqueue time. The worker resolves all six axes before it builds either prompt. Two prayers on one order get two different seeds, so a Second Prayer buyer never receives the same photo twice.

Member jobs do not use the angle axis. A monthly member delivery takes its scene from scene_variant = (member_no + month_index) mod n_variants and its treatment from (member_no + year_offset) mod 4 (§9). The worker branches on the job type, not on the seed: angle = (variant_seed >> 4) mod n_variants is the per-order path only.

1.3 Worker call order

load BrandConfig → load donor ref for location_key → resolve 6 axes from variant_seed
  → render the printed card at @prayer-engine/card (no model call, no cost)
  → Pass 1: donor ref + base scene + [TIME_OF_DAY] + [LIGHT] + [ANGLE] + CONSTRAINTS
  → Pass 2: edit on pass-1 output, the rendered card attached as a second image,
            hand variant + card variant + [SIDE] + [CARD_DISTANCE]
  → QA read-back: one vision call over BOTH frames, six checks (Section 4.5)
  → upload → deliverables row

The Pass-1 frame is held until QA finishes, because three of the six checks compare the two frames against each other. QA failure on any check re-runs Pass 2 only. The Pass-1 scene is kept. Three Pass-2 failures escalate the job to the dead-letter queue and set prayers.status = 'failed' (OPS §5).

A bump order renders the card once and attaches it at both locations, so the two photographs cannot disagree about the card. Before 2026-09-28 the front end attached a blank card and the bump attached the front end's finished photograph and was told to copy the writing out of it; that hand-off existed only because two independent hand-written renders never matched, and it is gone.


2. DONOR REFERENCE SPEC (per location)

One donor image per location key. Sourced by hand on Pinterest, approved once, then frozen. Replacing a donor changes the look of every future photo at that site, so a replacement needs a fresh sign-off line.

2.1 Shared acceptance checklist

Every donor must pass all eight before sign-off:

  1. Resolution: at least 2000 px on the long edge. No upscaled and no AI-generated source.
  2. No watermark, no stock-agency overlay, no photographer signature, no border, no collage.
  3. No crowd. Zero recognisable faces. A distant blurred figure is acceptable; a group of pilgrims is not.
  4. No text in frame: no signage, no plaques with readable lettering, no banners, no burned-in captions.
  5. Matches the scene variants listed for that location in Section 6. If the donor shows a part of the site none of the variants describe, either the donor or the variants are wrong.
  6. Clean single viewpoint. One coherent perspective, not a panorama stitch, not a fisheye, not a tilt-shift miniature.
  7. Natural colour. No heavy HDR halos, no filter, no black-and-white, no teal-and-orange grade.
  8. Foreground surface present. A ledge, rail, stone, floor or plaza surface in the lower third that the Pass-2 hand can plausibly enter in front of.

2.2 Shared Pinterest search recipe

Run all five queries for the location. Pin 15 to 25 candidates, then filter with the checklist above. Sort by Pinterest's default relevance, never by "most saved": most-saved skews to heavily edited travel-influencer frames that fail items 7 and 2.

2.3 Sign-off line

Record one line per approved donor in brands/<slug>/assets/donor/APPROVED.txt:

DONOR APPROVED | location_key=<key> | file=brands/<slug>/assets/donor/<key>.jpg
  | long_edge=<px> | source_url=<pinterest pin url> | approved_by=<name> | date=<YYYY-MM-DD>

A location with no APPROVED line does not go into production. The worker refuses the job.

2.4 st_peters_confessio (v1)

Must show: the basilica interior looking down the nave toward the baldachin and main altar, or the marble altar rail in the near foreground with the nave behind. Camera at standing eye height, 15 to 40 m from the altar. Travertine, polished coloured marble, gilded bronze and coffered vault all visible. Daylight from the clerestory windows, warm, no flash, no night-only frame.

Queries: st peters basilica interior nave empty · st peters basilica baldachin altar light shafts · vatican basilica marble altar rail close up · st peters basilica interior morning light no people · saint peters rome nave columns coffered ceiling

2.5 western_wall (Wave 1)

Must show: the Wall's lower courses filling the frame, close enough that individual limestone blocks read as blocks and the folded paper notes in the seams are visible. Camera 1 to 3 m from the stone, straight on or slightly angled. Late-afternoon or golden-hour raking light on the stone texture. No prayer books, no chairs, no partition screens, no people at the Wall.

Queries: western wall jerusalem stones notes close up · kotel limestone blocks paper notes cracks · western wall golden hour stone texture · wailing wall jerusalem detail no people · kotel stones prayer notes macro

2.6 fatima (Wave 1 expansion)

Must show: the Chapel of the Apparitions (Capelinha das Aparições) open-sided shelter with its plain columns and the marble pillar marking the apparition site, or the adjacent candle-burning area with its long steel troughs and racks of tall white candles. Camera at standing height, 5 to 20 m out. Overcast or warm evening light, the open plaza visible beyond.

Queries: capelinha das aparicoes fatima chapel apparitions · fatima portugal candle burning area candles · fatima sanctuary chapel of apparitions evening · fatima shrine plaza empty architecture · fatima candles wax troughs candles rack

2.7 lourdes

Must show: the grotto rock face with its arched natural niche and the statue recess, candles in glass holders massed at the base, wet stone, the spring channel. Camera 3 to 10 m out, slightly below the niche looking up, or level. Overcast daylight or evening candle glow.

Queries: lourdes grotto rock niche candles france · grotte de lourdes massabielle rock face · lourdes candles glass holders burning · lourdes grotto wet rock spring water close · lourdes sanctuary grotto no crowd

2.8 holy_sepulchre

Must show: the Stone of Anointing slab with the hanging brass oil lamps above it, or a candlelit stone-arched chapel interior in the same church. Camera 2 to 6 m out at standing height. Warm lamp light, incense haze acceptable, no flash.

Queries: church of the holy sepulchre stone of anointing · holy sepulchre jerusalem oil lamps brass hanging · holy sepulchre interior candles stone arches · jerusalem holy sepulchre chapel lamplight empty · stone of anointing slab close up lamps

2.9 guadalupe

Must show: the tilma of Our Lady of Guadalupe in its gold frame above the altar of the Nueva Basílica, framed straight on from the nave floor, or the modern basilica interior with its curved timber ceiling and radiating pews. Camera 10 to 30 m out. Even interior daylight through the clerestory band.

Queries: basilica de guadalupe interior tilma altar · nueva basilica guadalupe mexico interior ceiling · virgen de guadalupe tilma gold frame altar · basilica guadalupe nave pews architecture empty · guadalupe basilica interior wide no people


3. PASS 1 PROMPT TEMPLATES (reference-conditioned)

Pass 1 always runs with the donor image attached as a reference. The prompt opens with the match instruction, then the base scene, then the three randomized blocks, then the constraints block.

3.1 Prompt assembly order

[MATCH INSTRUCTION] + [BASE SCENE for location_key] + [TIME_OF_DAY block]
  + [LIGHT block] + [ANGLE block from Section 6] + [PASS-1 CONSTRAINTS]

3.2 Match instruction (identical for every location)

Using the attached reference photograph as the authority for this place: match the
composition, materials and light of the reference image. Keep the same architecture,
the same stone and metal surfaces, the same proportions and the same spatial depth.
This is the same real location photographed again on a different day, not an
interpretation of it. Re-render it as a new photograph under the conditions below.

3.3 Base scenes (one per location key)

st_peters_confessio

Photorealistic interior of St. Peter's Basilica in Rome. Polished marble columns and pilasters
receding down the nave toward the gilded baldachin and the main altar, coffered and gilded vaulting
overhead, pale travertine underfoot, broad marble ledges and balustrades running along the altar steps.

western_wall

Photorealistic close view of the Western Wall in Jerusalem. Massive ancient limestone blocks in warm
honey tones, weathered surfaces cracked and pocked with centuries of wear, narrow folded paper notes
pressed into the seams, the wall continuing upward out of frame, worn plaza stones at its base.

fatima

Photorealistic view of the Chapel of the Apparitions at the Sanctuary of Fatima in Portugal. The
plain open-sided shelter on slender columns, the marble pillar marking the apparition site, pale
worn stone floor, the wide open plaza beyond, rows of tall white candles burning in steel troughs.

lourdes

Photorealistic view of the Lourdes Grotto in France. The natural rock niche with its shallow arched
hollow, dozens of candles in glass holders massed and glowing at the base of the rock, clear spring
water running along a stone channel, damp rough rock face, pale flowers on the wet stone ledges.

holy_sepulchre

Photorealistic view of the Stone of Anointing inside the Church of the Holy Sepulchre in Jerusalem.
The broad pale limestone slab catching the glow of hanging brass oil lamps on chains, rough stone
walls and aged pillars around the chapel, candle niches in recessed arches, incense haze in the air.

guadalupe

Photorealistic interior of the Basilica of Our Lady of Guadalupe in Mexico City. The curved timber
ceiling sweeping overhead, the broad radiating nave, the gold-framed tilma of Our Lady raised above
the altar and lit warm against the dark sanctuary wall, even daylight from the high clerestory band.

3.4 Time-of-day blocks

MORNING:   early morning, cool clear air warming at the edges, long low light entering
           from one side, the space still and freshly lit.
AFTERNOON: full afternoon daylight, bright and even, warm highlights on stone, short
           soft shadows, the space at its most open and legible.
EVENING:   late evening, deep amber daylight nearly gone, flame and lamp light taking
           over as the dominant source, long warm shadows.
NIGHT:     night, no daylight at all, only candle, lamp and fixture light, deep warm
           pools against cool dark stone, high contrast, intimate and hushed.

3.5 Light blocks

SOFT DIFFUSE:       soft diffuse light with no hard edges, gentle falloff, even exposure across the
                    frame, calm and open.
RAKING DIRECTIONAL: hard directional light raking across the surfaces at a low angle, picking out
                    every grain and chisel mark, visible light shafts and dust motes in the air.
POOLED FLAME:       light pooled from flames, warm amber falling off fast into shadow, every flame a
                    distinct source, the rest of the frame sinking into warm dark.

3.6 Pass-1 constraints block (append to every Pass 1)

Photorealistic, 35mm documentary photograph, natural colour, shallow depth of field,
reverent mood. Colour: neutral, true-to-life white balance, not orange, not warm-filtered,
even in evening and night frames; the marble reads white and grey, the bronze reads dark
brown, flame light stays small and local. A candid photograph taken on a phone, not a
studio image. Portrait orientation, 3:4. No people, no faces, no hands, no readable text
anywhere, no signage, no plaques, no logos, no watermark, no borders, no distortions, no
duplicated architecture.

4. PASS 2 CARD-IN-HAND RECIPE

Pass 2 is an edit on the Pass-1 frame. It puts one hand holding the buyer's printed card into the foreground and changes nothing else about the scene.

The card is an input. Pass 2 renders no text. @prayer-engine/card has already typeset the buyer's name and prayer onto it, and the finished PNG is attached as the SECOND image with the ## CARD REFERENCE clause. There is no {FIRST_NAME}, no {PRAYER_TEXT} and no {SPELLED_NAME} in this prompt — a model asked to write a name is a model that can get it wrong, and the whole point of printing the card first is that nothing on it is left to a model.

The card is HELD. A held card has no surface. surface, ledge, plinth, step, table and stand appear nowhere in this recipe outside a negative clause — see §4.1b for what happens when one of them does.

One clause, at every location. [CARD_REFERENCE] renders on the front end and on the bump alike, because the same rendered card is attached to both.

4.1 Edit instruction template

Both passes run on the newest GPT image model (openai/gpt-5.4-image-2 via OpenRouter). Shaw's call 2026-09-16: GPT only for image generation, no Gemini. The vision read-back (§4.5) is a separate reading model and generates nothing.

Edit this photograph. Keep the place exactly as it is: the same architecture, marble,
bronze, surfaces, light, time of day and camera position. A person's hand enters the frame
from the lower [SIDE] foreground, close to the camera, holding up a [CARD_VARIANT] so the
card faces the camera squarely, [CARD_DISTANCE] [HAND_VARIANT]. The fingers grip the card
at its lower edge and along one side, thumb in front, fingertips visible, no finger
crossing the printed text.

[CARD_REFERENCE]

The card is lit by the same light as the scene behind it, with a soft natural shadow where
the fingers press it. The card and everything printed on it are the sharpest thing in the
photograph; the scene behind stays in focus but a little softer.

Colour: neutral, true-to-life white balance, not orange, not warm-filtered; the marble
reads white and grey, the bronze reads dark brown. This is a candid photograph taken on a
phone by the person holding the card, not a studio image. Portrait orientation, 3:4. Add
nothing else beyond what is stated above. No other text anywhere in the frame, no signage,
no logos, no watermark, no second hand, no face.

Slots: [SIDE] is left or right, from side = (variant_seed >> 11) mod 2 (§1.2). [CARD_VARIANT] is the E1 / E2 / E3 card description from Section 7. [CARD_DISTANCE] follows the prayer length (4.2 rule 2; blocks in 4.3a). [HAND_VARIANT] is one of the three blocks in 4.3: one persona hand, three small variations. [CARD_REFERENCE] is the §7 second-image clause, and it renders at every location.

There is no {FIRST_NAME} and no {PRAYER_TEXT}. The buyer's words are on the attached card, printed by @prayer-engine/card before this prompt is built. A template with a name slot in it is asking a model to spell a stranger's name correctly, which is a thing models get wrong and a thing we no longer have to risk.

4.1a The bump is a second LOCATION (The Apostle's Tomb, $9.97)

The Wider View is dead. It was a second photograph of the same placement, cropped wider — a second generative act off one Pass-1 frame, which meant re-solving every realism rule per order, and it asked the buyer to pay to see less of her own prayer. Replaced 2026-09-21 by a second location: the same card, carrying the same prayer, photographed in the Vatican Grottoes — the crypt three metres beneath the basilica floor. The same card, one place further in.

There is no candle clause. Removed 2026-09-21, with the bump it served. Every order used to be promised a candle lit beside the card, and the promise could not be kept: a visitor cannot light a votive at the papal altar, and the votive lights that do exist at St. Peter's are electric (REALISM-RULES R1). Pass 2 places no light of any kind, and no fittings or props — no stand, no candle, no lamp. It puts the card into the light the place already has.

Two locations, two Pass-1 frames, one gate. The job payload's bump_taken boolean — true when the buyer took The Apostle's Tomb at checkout — makes the worker run the whole pipeline a second time at the bump location: its own Pass-1 frame from the Grottoes donor, then its own Pass-2 edit on it. [WIDE_VARIANT] and the wide composition no longer exist; a pack ships no such section and the token is gone from the edit instruction.

The front-end photograph is the base deliverable. The bump is a second deliverable of the same kind, and it runs the same six QA checks. The card is fully legible at the crypt, so nothing is relaxed the way the old wide gate relaxed it.

The pair must never disagree about the visit. Two places cannot share one frame, so what they share is everything else: both renders read the same variant_seed for the hand, the card and the side, so the same hand holds the card in both, and the light quality and crowd density match. REALISM-RULES R8 is the contract — the pair must look like one trip, and the bump must not be confusable with another basilica shot. The crypt is worth $9.97 only because it is visibly the place under the place.

REALISM-RULES applies to every frame. Nothing invented, no fittings or props, nothing on the floor, the camera at standing eye level, the card never the brightest object. Read prayer/REALISM-RULES.md before touching anything in this section.

4.1b The resting surface is a PASS-1 property — and the frame must NOT have one

Inverted 2026-09-21. This section used to require a raised resting surface in every Pass-1 frame. It now forbids one. The deliverable is a card held up in the hand, and a held card has no surface.

The original finding stands, and it is the whole reason this is a Pass-1 section rather than a Pass-2 instruction. Found 2026-09-21: a scene failed the "card never on the floor" rule twice in three attempts while an almost identical scene passed every time. The Pass-2 wording was not the difference — the base frame was. The failing Pass-1 frame had no raised surface anywhere in the near foreground, only open marble paving, so Pass 2 set the card down on the floor because that was the only surface it had.

A thing the frame does not contain gets invented by the model. That is why the fix lives in Pass 1, and it is why removing the surface is also a Pass-1 job:

  1. No Pass-1 frame may offer up a flat raised surface in the near foreground — no ledge, plinth, step, balustrade top, stand or table with an empty top for a card to rest on. The lower third of the frame is the floor and the architecture, and it offers nothing up.
  2. The words are the problem, not only the meaning. surface, ledge, plinth, step, table and stand may appear in a Pass-1 or Pass-2 prompt only inside a clause that forbids them. The shipped bad photograph came from a Pass-2 prompt that asked for a card held in the air and for a lit candle "standing on the surface just beyond the card's lower edge". The model could not satisfy both, invented the surface, and filled it with candles.
  3. Every angle entry in every location pack in Section 6 must satisfy this. An angle whose near foreground is a ledge with an empty top is not a usable angle. The real architecture — the Confessio balustrade, a pier base, the inlaid floor — stays in frame; what must not happen is a surface offered up for the card, or one invented near it.
  4. When a frame keeps coming back with the card resting on something, regenerate Pass 1. Do not keep rewriting Pass 2.

See prayer/REALISM-RULES.md R2, R2a and R2b.

4.2 Legibility rules

The card is printed, so the old rules about making a model write neatly are gone. What survives is everything about the photograph — how big the card is in frame, what may cross it, how it is lit — because that is what the buyer sees and what the QA read-back has to be able to read.

  1. Cap is 320 characters, enforced at the survey, again at /api/offer, and again by the renderer. Never truncate at render time. A stored prayer over the cap is a funnel bug, not a prompt problem: fail the job loudly.
  2. Card distance follows prayer length ([CARD_DISTANCE], §4.3a): up to 120 characters FAR, 121 to 180 MID, 181 and above NEAR. FAR is the norm at both locations — the card at about a third of the frame with the place rising behind it. MID and NEAR exist only so a long prayer still reads, and half the frame width is a hard ceiling: past it the place stops being the subject (R4).
  3. Nothing occludes the printed text. The grip is at the lower edge and one side, thumb in front, and the prompt says so. A finger over a word is a QA failure, not a stylistic choice.
  4. The card and its printed text are the sharpest element. State it. The model otherwise focuses on the architecture it just rendered — and these are finely detailed frames, so its instinct is strong.
  5. Colour is neutral. State the white balance every time: marble white and grey, bronze dark brown, not orange. The model otherwise warms the scene toward candlelight orange, and the paper goes cream with it.
  6. The text is reproduced, never re-set. The clause says so explicitly: same words, same line breaks, same typeface, same size, same position. A model that "tidies" the typography has changed the buyer's card, and the QA gate compares the read-back to the source word by word.
  7. A retry changes the grip and the card slot, not the {FIRST_NAME} spelling. There is no name to spell any more, so a second attempt steps the hand and the card variant on instead of asking the same question again (varyAxesForRetry). The scene, light, angle and side are held — they are what the order promised.

4.3 Hand variants

One hand per brand: the persona's. The three entries are the same hand with small differences so the set does not repeat exactly. Never a ring, never a watch, never a face.

H1: The hand is a man's in his fifties, light warm skin, clean short nails, faint lines
    at the knuckles, no ring, no watch, the cuff of a plain black sleeve just visible at
    the wrist.
H2: The hand is a man's in his fifties, light warm skin, clean short nails, faint lines
    at the knuckles, no ring, no watch, the plain black sleeve pushed a little further
    back so more wrist shows.
H3: The hand is a man's in his fifties, light warm skin, clean short nails, faint lines
    at the knuckles, no ring, no watch, the cuff of a plain black sleeve at the wrist,
    the thumb pressing a touch harder so the card dips slightly.

4.3a Card distance

Selected by the prayer length in characters: up to 80 = FAR, 81 to 150 = MID, 151 and above = NEAR.

FAR:  held at arm's length so the card fills about one third of the frame width; the place
      stays the subject of the photograph and the card is the detail in it
MID:  held a little closer so the card fills about two fifths of the frame width, slightly
      tilted, every line easy to read
NEAR: held closer still so the card fills JUST UNDER half the frame width and every line is
      clearly readable, the place softening behind it. HALF IS A HARD CEILING: a card wider
      than half the frame is a failure, so stay under it and leave the place the subject

4.4 Filled examples

Both examples below are the prompt only. The card they refer to is a rendered file, and it is attached to the request as the second image — nothing about the card's appearance is written out in the prompt, because writing it out is what used to make every render a different card.

Example 1, short prayer. st_peters_confessio, morning, hand H1, card E1, distance FAR. The attached card carries the name Maria and the prayer Heal my mother, and give her strength for the year ahead.

Edit this photograph. Keep the place exactly as it is: the same architecture, marble,
bronze, surfaces, light, time of day and camera position. A person's hand enters the frame
from the lower left foreground, close to the camera, holding up the buyer's printed prayer
card from the attached image, crisp and flat with clean square corners, reproduced exactly
as attached, the printed header across the top unchanged so the card faces the camera
squarely, held at arm's length so the card fills about one third of the frame width; the
place stays the subject of the photograph and the card is the detail in it. The hand is a
man's in his fifties, light warm skin, clean short nails, faint lines at the knuckles, no
ring, no watch, the cuff of a plain black sleeve just visible at the wrist. The fingers
grip the card at its lower edge and along one side, thumb in front, fingertips visible, no
finger crossing the printed text.

A SECOND IMAGE is attached to this request. It is the buyer's printed prayer card — already
made, already printed, complete — and IT is the authority for the card in your output, not
your own idea of a prayer card. Reproduce the attached card exactly and hold it up in the
photograph: exactly the same size and the same portrait proportions, exactly the same plain
white card stock, exactly the same printed header in the same position at the top, and, on
the field below the rule, exactly the same printed text, word for word, line for line, in
the same typeface, the same size and the same position.

EVERY WORD ON THIS CARD IS PRINTED TYPESET TEXT. Nothing on it is handwriting, nothing on it
is your own lettering, and you are not writing anything: you are photographing a card that
already exists. Do not re-write, re-type, re-space, re-break, translate, correct, add to,
shorten, summarise or re-arrange any of the text. Do not change a single character of it. Do
not add a date, a place, a signature, a number, a URL or any other printed mark anywhere on
the card. Reproduce the card as it is, and reproduce the whole of it.

The card is lit by the same light as the scene behind it, with a soft natural shadow where
the fingers press it. The card and everything printed on it are the sharpest thing in the
photograph; the scene behind stays in focus but a little softer.

Colour: neutral, true-to-life white balance, not orange, not warm-filtered; the marble
reads white and grey, the bronze reads dark brown. This is a candid photograph taken on a
phone by the person holding the card, not a studio image. Portrait orientation, 3:4. Add
nothing else beyond what is stated above. No other text anywhere in the frame, no signage,
no logos, no watermark, no second hand, no face.

Example 2, full prayer at the cap. western_wall, evening, hand H2, card E2, distance NEAR. Only the distance clause and the card slot change; everything else in example 1 is identical. This is what the 320-character cap renders as on a card whose text runs seven lines.

...held closer still so the card fills JUST UNDER half the frame width and every line is
clearly readable, the place softening behind it. HALF IS A HARD CEILING: a card wider than
half the frame is a failure, so stay under it and leave the place the subject...

4.5 QA read-back (run on every Pass-2 output before upload)

This is a machine check, not an eyeball. The worker sends two images in one call — the Pass-1 frame and the Pass-2 output — to a vision model through OpenRouter (OPENROUTER_API_KEY, default google/gemini-3.5-flash) with one instruction: transcribe the card face exactly as printed and answer four scene questions as booleans. It then compares the transcription against the source strings.

The source strings are ours, not a model's. Since 2026-09-28 the pipeline renders the card itself, so checks 1 and 2 grade the photograph against text we authored and printed, rather than against a model's attempt at handwriting. What they catch is a card that came back unreadable — a word under a finger, ink blown out by a highlight, a card the model restyled — not a misspelling. The thresholds are unchanged because the failure mode is unchanged in kind: an unreadable card is not deliverable. The mechanism, cost and retry budget are specified in OPS-RUNBOOK §5 step 6; this section is the prompt-side contract.

Six gating checks. All six must pass. Each one names the field of the call that verifies it.

# Check Verified by Pass condition
1 Name legible card_text The read-back name matches the source first name exactly, case-insensitive. Not a spell-check of a model's handwriting — a check that the PRINT is legible and that no finger or highlight has swallowed a letter.
2 Prayer legible card_text At least 95% of the source prayer's words appear in the read-back, in source order. No dropped clause, no line the card's own curve put out of view.
3 Same place same_place The Pass-2 image shows the same real location as the Pass-1 image: recognisably the same architecture and surfaces from a similar viewpoint, allowing for the added hand and card and a modest change of framing. False when the room, building or architecture was replaced. (The Pass-2 model re-renders the place rather than editing pixels, so pixel-level sameness is not the test. Changed 2026-09-16.)
4 No new text new_text Nothing readable in the Pass-2 frame except the card face AND text that was already visible in the Pass-1 frame (the basilica's own inscriptions and friezes do not count). Fails on a second card, a watermark, a caption or lettering that appeared only in Pass 2. (Changed 2026-09-16: the frieze inscription failed every St. Peter's frame under the old rule.)
5 Card not occluded card_unoccluded No finger, thumb, flame or architectural edge crosses the printed text on the card.
6 Hand plausible hand_plausible One hand with five fingers, one card, no face, no second person.

The call. One request, two images attached in order (Pass 1 first, Pass 2 second), one JSON response:

{ "card_text": "<the card face, transcribed exactly as printed>",
  "same_place": <bool>, "new_text": <bool>,
  "card_unoccluded": <bool>, "hand_plausible": <bool> }

Checks 1 and 2 are computed by the worker from card_text. Checks 3 to 6 read same_place, new_text, card_unoccluded and hand_plausible straight. The candle-count check is gone (2026-09-21) along with the bump it proved; it must not be reinstated.

On any failure, re-run Pass 2 only — the Pass-1 scene is kept — using the name-spelling recovery prompt from 4.2 rule 9. Maximum three Pass-2 attempts. After the third, the job dead-letters and prayers.status goes to failed. The two-image read-back costs about $0.02 per attempt — one call, six checks — which is the line carried in SPEC-PRAYER-ENGINE §3 and ECONOMICS §2.


5. EXAMPLE-DELIVERABLE HERO SET

Seven hero deliverables per brand, because the copy needs seven distinct images: 1 landing hero, 3 on /create/ready, 3 on /offer. That is exactly what BrandConfig.location.examplePhotos declares (// 7 paths) and what both launch checklists gate on. Fake-name stubs only. Generated once, approved by hand, shipped as static assets. They are not regenerated per deploy.

Build each hero by running the standard Pass 1 and Pass 2 with the values in the table. Hero H-1 is written out as the worked model; heroes H-2 to H-7 are the same two prompts with the table's values substituted. Every hero is the standard deliverable: no bump (bump_taken false). Each one shows the card in the place's own light under REALISM-RULES.

5.1 st_peters_confessio hero set

# Name Prayer text on card Time of day Angle Hand Card Page
H-1 Maria "Heal my mother, and give her strength for the year ahead." morning A2 close altar rail H1 E1 landing hero
H-2 Denise "For my husband's surgery on the 14th. Steady the hands of his doctors and bring him home to us." afternoon A1 wide nave H3 E2 transition
H-3 Robert "Thank you for forty-one years with her. Watch over her now." evening A4 ledge close H2 E1 transition
H-4 Kathleen "For my daughter on the night shift. Bring her home safe every morning." night A5 low-angle columns H1 E3 transition
H-5 Maria "Let my son find work this month, and let him believe in himself again while he looks." morning A3 over the rail H1 E3 sales page
H-6 Denise "For peace in this house, and for the strength to keep asking." evening A2 close altar rail H3 E2 sales page
H-7 Robert "For my brother, nine years gone this week, and still not a day without him." afternoon A4 ledge close H2 E1 sales page

H-1 Pass 1, with brands/st-peters/assets/donor/st_peters_confessio.jpg attached. Assemble the 3.2 match instruction, then the st_peters_confessio base scene from 3.3, then these three blocks, then the 3.6 constraints block:

[ANGLE A2]  Framed close on the polished marble altar rail: the veined stone ledge crosses the lower
            foreground, candle stands just beyond it, the nave and altar softly blurred behind.
[MORNING]   Early morning, cool clear air warming at the edges, long low light entering from one
            side, the space still and freshly lit.
[RAKING]    Hard directional light raking across the surfaces at a low angle, picking out every
            grain and chisel mark, visible light shafts and dust motes in the air.

H-1 Pass 2: the 4.1 template with [SIDE]=left, [CARD_VARIANT]=E1 crisp, [CARD_DISTANCE]=FAR, [HAND_VARIANT]=H1, and the row's card rendered by @prayer-engine/card from the name Maria and the prayer Heal my mother, and give her strength for the year ahead.

5.2 western_wall hero set

The same seven heroes, same names, same prayer texts, re-shot at the Wall. Swap the base scene to western_wall, the donor to brands/<slug>/assets/donor/western_wall.jpg, and the angles to the 6.4 list.

# Name Time of day Angle Hand Card Page
H-1 Maria morning D1 notes between stones H1 E1 landing hero
H-2 Denise afternoon D2 wall wide H3 E2 transition
H-3 Robert evening D1 notes between stones H2 E1 transition
H-4 Kathleen night D3 low-angle stone mass H1 E3 transition
H-5 Maria morning D3 low-angle stone mass H1 E3 sales page
H-6 Denise evening D4 seam close with note H3 E2 sales page
H-7 Robert afternoon D2 wall wide H2 E1 sales page

For western_wall, [SIDE] alternates the same way, but the Pass-2 surface language changes one clause. The hand holds the card up against the stone, not over a rail. Replace the first sentence's tail with:

...holding up a small [CARD_VARIANT] prayer card flat against the limestone face, a
hand's width from the folded paper notes in the seams, so the card faces the camera
squarely and fills roughly one third of the frame width.

6. LOCATION PACKS (angle and scene vocabulary)

The [ANGLE] block for Pass 1. One pack per location key, matching the donor specs in §2.4 to §2.9 one for one. Append the 3.6 constraints block to every one.

6.1 st_peters_confessio

A1 WIDE NAVE: shot from floor level at the entrance to the nave, an immense symmetrical corridor of
   marble columns and gilded vaults converging on the distant lamplit altar, the floor's geometric
   stonework leading the eye straight to the glow.
A2 CLOSE ALTAR RAIL: framed close on the polished marble altar rail, the veined stone ledge crossing
   the lower foreground, the marble balustrade just beyond it, the nave and altar softly blurred behind.
A3 OVER THE RAIL: framed from just behind and above the altar rail, the rail's edge cutting across
   the lower foreground, the altar under its dark bronze baldachin and the receding nave beyond it gently out of focus.
A4 LEDGE CLOSE: framed on a broad marble ledge, balustrade or pier base in the near foreground, its
   empty polished top running across the lower third, the nave and gilded vaulting soft beyond it.
   This is the resting surface a card is later set down on: real architecture, no fittings, no props.
A5 LOW-ANGLE COLUMNS: looking up along a colossal marble column toward the coffered and gilded
   vault, light raking across the carved capital, the altar small and distant below.

6.2 holy_sepulchre

B1 STONE SLAB CLOSE: the worn pale limestone slab of the Stone of Anointing filling the
   frame, brass oil lamps glowing just above, warm reflections pooling on the stone.
B2 OIL-LAMP ROW: a row of hanging brass oil lamps on chains, small flames steady inside
   glass chimneys, the pale slab glowing below, aged walls receding into warm shadow.
B3 CHAPEL ARCHES WIDE: the candlelit chapel interior, low stone arches and worn pillar
   capitals, clusters of small candles on ledges against shadowed stone, incense haze.
B4 STONE LEDGE CLOSE: a stone ledge beside a lit votive candle in the near foreground,
   lamplight pooling across it, the slab and blurred arches glowing behind.

6.3 lourdes

C1 GROTTO NICHE: the arched natural rock hollow centred in frame, candles in glass holders
   glowing amber at its base, the niche lit against darker rough rock, flowers on the ledge.
C2 CANDLE GLASS MACRO: candles in glass votive holders close, flames sharp in their cups,
   wax light spilling across wet rock, cool cave bokeh behind.
C3 WATER AND FLOWERS: low along the stone channel, clear spring water running over smooth
   wet stone catching the candle flicker, pale flowers among the rocks at the water's edge.
C4 ROCK LEDGE CLOSE: a flat rock ledge in the near foreground among small white flowers, a
   candle in a glass holder burning beside it, the niche arch blurred in amber behind.

6.4 western_wall

D1 NOTES BETWEEN STONES: close on a cluster of small folded paper notes pressed into the gaps
   between the limestone blocks, edges curled by weather, light raking across the pitted stone.
D2 WALL WIDE: the enormous wall spanning the frame, low sun bathing the lower courses in gold
   while the upper stones fall into soft shadow, the empty plaza stones in front.
D3 LOW-ANGLE STONE MASS: looking up the towering face of gigantic limestone blocks from the plaza
   floor, chisel marks and erosion pits sharp in raking light, notes in the seams.
D4 SEAM CLOSE WITH NOTE: framed tight on a single seam between two massive blocks with one folded
   note tucked upright in it, the courses above and below falling into golden bokeh.

6.5 fatima

Base scene in 3.3. Three variants:

F1 CHAPEL OF THE APPARITIONS: the open-sided shelter framed head on from the plaza, slender columns
   holding the low roof, the marble apparition pillar centred beneath it, worn pale stone floor.
F2 CANDLE-BURNING AREA: the long steel troughs of the candle area, racks of tall white candles in
   dense rows, wax pooling down the trough walls, heat shimmer, the chapel pale in the background.
F3 CANDLE ROW CLOSE: close along one rack of tall white candles, flames sharp and leaning together
   in the air, melted wax sheeting down the steel, the open plaza a bright blur beyond the flames.

6.6 guadalupe

Base scene in 3.3. Three variants:

G1 TILMA ABOVE THE ALTAR: framed straight on from the nave floor, the gold-framed tilma of Our Lady
   raised above the altar, lit warm against the dark sanctuary wall, the nave soft below it.
G2 BASILICA INTERIOR WIDE: the modern basilica interior from the rear of the nave, curved timber
   ceiling sweeping overhead, radiating pews converging on the sanctuary, even clerestory daylight.
G3 SANCTUARY CLOSE: close on the sanctuary rail and altar below the tilma, polished stone and pale
   timber in the near foreground, votive candles to one side, the framed tilma gold and blurred.

7. CARD ART (E1 to E3)

The card is a printed artefact, not a description and not a generation. Shaw's directive, 2026-09-28: "The card image needs to be the same card. The size of the card needs to be exact and the exact same size each time… the text is always printed. It's always going to be uniform, with a character limit."

The card is rendered by code, before Pass 2 runs, from the buyer's own strings:

Property Value
Format 105 x 148 mm portrait (A6), 3 mm bleed, rounded corner
Render 1240 x 1748 px at 300 dpi
Stock plain white, flat fill
Header our standalone lockup, deep maroon #7B1E22, 52% of the width at 7.5% of the height
Rule 46% of the width, 5 px, with one centred diamond, 3% of the height below the lockup
Footer the emblem alone at 89.5%, so a crop that cuts the header still reads as our card
Name EB Garamond SemiBold, maroon, one line, 168 px (the fit ladder steps down only for an outlier)
Prayer EB Garamond Regular, near-black, wrapped to the field, 54 px, one size for every order (a guard rung at 46 px exists only for a pathological single-glyph string)
Field 10% margin each side, 24.5% to 84.5% of the height

The whole geometry lives in prayer-engine/packages/card/layout.json, and the renderer, the asset builder and the TypeScript layout all read that one file so the card cannot drift from its own description. CARD-SPEC.md is the authority for the layout; this section is the prompt pack's view of it.

It is attached, never described. The rendered card is a second input image, so the printed header and the printed text are the same on every order instead of something the model invents afresh each render. The clause is ## CARD REFERENCE in the st-peters pack, and it renders at both locations. A model told to "add a logo" draws a different one every time — the same failure class as the invented tear below.

What is left to describe is the stock and the hold, which is [CARD_VARIANT]. The card's size, proportions, colour, header and text are invariant and are never a variant.

E1 CRISP: the buyer's printed prayer card from the attached image, crisp and flat with clean
   square corners, reproduced exactly as attached, the printed header across the top unchanged.
E2 POCKET: the buyer's printed prayer card from the attached image, its corners very slightly
   softened from being carried in a pocket, the card itself and everything printed on it unchanged.
E3 CURVED: the buyer's printed prayer card from the attached image, held with a slight natural
   curve so the paper bows a little, the card itself and everything printed on it unchanged.

The card's text rule is unchanged and absolute: every word on the card is printed typeset text, and it is reproduced exactly — same words, same line breaks, same face, same size, same position. Never handwriting, never re-spaced, never re-broken, never re-typed. The model's only freedom is the hand, the stock and the angle of the hold.

Nothing else is printed on it. No date (the delivery window varies), no place name (the same card goes to two locations), no signature, no QR code, no serial, no website, no price. The reasons are in CARD-SPEC.md.

Standalone card shot (no scene, no hand) for thumbnails and ad stills. The card is rendered first and attached, exactly as in the main path:

[the buyer's rendered card, attached as a second image], photographed straight on under soft
even daylight on a plain neutral surface, centred composition, gentle shadow. Reproduce the
attached card exactly — the same size and proportions, the same printed header and rule, and the
same printed name and text, word for word and line for line. Everything on it is printed typeset
text; do not re-write, re-break, add to or re-arrange any of it. No other text, no watermark,
neutral colour, photorealistic.

7.1 The card is a code asset with a fixed geometry

There is no card MASTER PNG to approve any more. The card's identity is layout.json plus the logo files plus one typeface, and the renderer turns those into a card per order. Change layout.json and every future card changes with it; a card rendered before the change stays as it was shipped, which is the honest record of what that buyer received.

packages/card/layout.json                the geometry — the single source of truth
packages/card/assets/EBGaramond[wght].ttf  the typeface, sliced to Regular + SemiBold
packages/card/assets/fonts/              the two sliced, subsetted statics
packages/card/src/generated.ts           the built assets: header, footer, glyph atlases
packages/card/scripts/build-assets.py    builds them; `--check` fails the build when stale
packages/card/scripts/render-sample.ts   renders samples to look at

Rebuild after any change to the layout, the logo or the font:

pnpm --filter @prayer-engine/card build:assets
pnpm --filter @prayer-engine/card render:sample /tmp/card-samples

brands/<slug>/assets/card/ no longer exists, and push-brand-assets.sh no longer uploads a template: the worker does not read the card from storage, it renders it.

8. PERSONA PORTRAIT PACK

Persona: a warm, believable priest in his late 50s. Silver hair, kind eyes, black cassock with white collar, Roman setting. Reference-image workflow: generate the MASTER first, then drive all follow-ups with the master attached as reference. Generate once per brand, 6 to 10 frames, same face across the whole set.

8.1 Master reference prompt (generate first, keep as the character anchor)

Studio portrait photograph, front-facing, of a warm believable Catholic priest in his late
50s: silver hair neatly combed, kind expressive eyes, gentle natural wrinkles, light warm
smile lines, wearing a black cassock with a white roman collar. Neutral soft light, clean
seamless warm-grey background, sharp focus on the eyes, 85mm portrait lens look, shallow
depth of field, natural skin texture, photorealistic, no text, no logos, no watermark.
This is the master reference image for the character.

8.2 Follow-up poses

All prefixed with: same man as the reference image, same face, same silver hair, same black cassock with white collar, now:

P1 LETTER BY CANDLELIGHT: writing a letter by hand at a worn wooden desk lit by one beeswax candle,
   flame light carving soft shadows across his face, a brass crucifix on the desk edge, dark beyond.
P2 WALKING AT DAWN: a distant wide shot of him walking alone toward St. Peter's Basilica at dawn,
   cool blue air with the first gold light on the dome, long empty cobblestone approach.
P3 CANDLE IN THE NAVE: standing in the dim nave holding a single lit candle at chest height, flame
   light on his kind face, marble columns and gilded arches falling into soft shadow behind him.
P4 HANDS OVER A CARD: medium close at the marble altar rail, his hands gently folded over a small
   parchment prayer card on the ledge, head slightly bowed, candlelight crossing the stone.
P5 SPEAKING TO CAMERA: medium shot speaking gently to camera, warm engaged expression mid-sentence,
   soft key light from a window, the basilica interior softly blurred behind him, calm.
P6 HANDS ONLY, PLACING: close on his hands only, no face, an older man's hands in black cassock
   sleeves placing a parchment prayer card among lit votive candles, wax glow on his knuckles.
P7 READING AT A TABLE: seated at a long wooden table strewn with dozens of small parchment cards
   and opened letters, reading one intently, warm lamplight from above, no readable text.
P8 COLONNADE AT GOLDEN HOUR: standing in a grand curved stone colonnade at golden hour, low warm sun
   flaring between the columns behind him, cassock catching the amber light, wide low angle.

All frames take 35mm documentary look, photorealistic, no text, no watermark, except P1 and P5 which take 85mm portrait lens look. P5 is the direct-address creative frame; P1 and P7 are the workhorse stills.


9. MONTHLY ROTATION

Members get a new photo every month. Three axes decide it.

9.1 Theme is the calendar month, always

Theme never rotates off the calendar month. There is no cursor, no queue, no stored index and no per-member theme offset. A December delivery is the December theme for every member in every year. Only the treatment and the scene variant move. This is deliberate: the theme is what the member recognises in the subject line, and it must match the month they are living in.

Catholic calendar (v1, and every Christian brand):

month_index Month Theme
0 January Renewal
1 February Devotion
2 March St. Joseph
3 April Easter light
4 May Marian month
5 June Sacred Heart
6 July Pilgrimage season
7 August Assumption
8 September Autumn intentions
9 October Rosary
10 November All Souls candles
11 December Advent

Jewish calendar (western_wall): month_index runs September = 0 through August = 11, so index 0 lands on the Hebrew year's opening month. Each row names the Hebrew month that predominantly overlaps that Gregorian month.

month_index Gregorian Theme
0 September Tishrei / High Holidays
1 October Cheshvan
2 November Kislev / Chanukah
3 December Tevet
4 January Shevat / Tu BiShvat
5 February Adar / Purim
6 March Nisan / Pesach
7 April Iyar
8 May Sivan / Shavuot
9 June Tammuz
10 July Av / Tisha B'Av
11 August Elul

The theme supplies scene dressing only: seasonal flowers, Advent wreaths, All Souls candle counts, Marian blues, a shofar on the ledge in Tishrei, lilies at Pesach. It never changes the site, the composition rule or the card.

9.2 Treatment

treatment = (member_no + year_offset) mod 4

member_no is a serial assigned at signup. year_offset is whole years since member_since. Four treatments: (0) dawn, (1) midday shaft, (2) candlelit evening, (3) blue hour. Twelve themes by four treatments is 48 combinations, so a member does not repeat a theme-and-treatment pair inside four years. The formula is the state; nothing is stored.

DAWN:              cool pale dawn light, first soft gold rim on edges, empty hushed air.
MIDDAY SHAFT:      bright vertical light shafts through high openings, dust motes, high
                   warm-cool contrast.
CANDLELIT EVENING: full candle and lamp glow, deep warm shadows, intimate amber palette.
BLUE HOUR:         deep blue ambient dusk with flames as the only warm points, high
                   warm-cool contrast.

9.3 Scene variant

scene_variant = (member_no + month_index) mod n_variants

n_variants is the length of the active brand's angle list in Section 6. This advances by one per month, so the angle changes on every delivery even when the treatment repeats.

This formula replaces the per-order angle axis, it does not stack with it. The worker takes an explicit branch: an order job resolves the angle from (variant_seed >> 4) mod n_variants (§1.2); a member job resolves it from scene_variant here and ignores the seed's angle bits entirely. The branch is on job type, and it is the only place the two paths differ.

9.4 Card text on member deliveries

Same card, same Pass 2, same rules. A member job resolves both card strings from the subscriptions row, never from an order — a standalone /membership member has no order row at all (OPS §5).

  • the card's name is subscriptions.first_name.
  • the card's prayer is subscriptions.current_intention when the member submitted one for that month, and that month's default prayer (brands/<slug>/prompts/themes.md, ## DEFAULT INTENTIONS, one per month_index, written in the member's voice) when they did not. A brand without that section falls back to the theme name. Added 2026-09-29: the theme name alone was too thin for a paid card.

Both go to @prayer-engine/card, which prints them and hands the finished PNG to Pass 2. Same 320-character cap either way. The QA read-back (4.5) checks the printed name in the photograph against that same subscriptions.first_name.


10. AD CREATIVE STILLS

Seven prompt blocks for the silent B-roll ad format. Each is a visual frame plus an animation note for the 5-second motion pass.

Revised 2026-09-30 — the candle is out, so these frames had to lose it. Shaw: "we are not doing a candle anymore remove." Four of the seven blocks below depended on candle flame for their light or their subject (AD7, AD1, AD4, AD5). Every one is now lit by the building's own lamps, which is the fact the evidence pack already had and never used: the Confessio burns rows of oil lamps kept perpetually alight (Stage 2 §5). The ad set was advertising a candle the offer does not sell, and the gate caught it with a mechanical check, not a read.

Revised 2026-09-25. Two things were wrong with this section:

  1. AD2 and AD4 showed a card on a surface. "a small parchment prayer card resting on the marble altar rail" and "tucking a parchment prayer card into a seam" both contradicted R2a/R2b, which exist because a card on a surface is how the shipped bad photograph happened — the model invented candles to fill the surface it was told about. A card in this product is HELD, always. Ad creative is not exempt from a rule the deliverable obeys.
  2. "Parchment" is dead. The card has been our printed A4 stock since 2026-09-25, carrying the Holy Cards of St. Peter's lockup, wordmark and rule. An ad showing a parchment scrap advertises a product we do not ship.

Where a frame contains the card it is conditioned on assets/card/template.png so the printed header is ours, and it shows the buyer's name legible with the prayer below it soft and unreadable (apps/worker/scripts/review-set.py is the standing recipe for that treatment). Where a frame contains a hand it is Fr. Marco's, driven off the character sheet.

Append the §3.6 constraints block to every still except those that include hands on purpose (AD1, AD3, AD5, AD6).

AD7 CARD IN HAND: close-up of a hand holding the printed Holy Cards of St. Peter's card up in
    front of a lamp-lit basilica interior, the printed header at the top, the buyer's printed
    name legible below it and the prayer beneath it soft and unreadable. CONDITION ON the card
    template and the persona sheet.
    ANIMATION: 5s slow push-in onto the card face, the lamps steady behind.
AD1 LAMPS AT THE CONFESSIO: photorealistic macro of the rows of oil lamps burning at the
    Confessio, their gold light low and warm against dark marble, deep bokeh of blurred gilded
    arches behind, one lamp burning a little brighter than the rest. NO candles — these are the
    basilica's own lamps.
    ANIMATION: 5s slow push-in toward the brightest lamp, no camera shake.
AD3 LIGHT SHAFTS: photorealistic wide interior of a great basilica at morning, parallel light shafts
    through high clerestory windows onto the stone floor, dust motes suspended in the beams.
    ANIMATION: 5s subtle sideways parallax drift, dust motes drifting through the shafts.
AD5 WIDE REVERENT SPACE: photorealistic ultra-wide of St. Peter's nave, symmetrical columns
    receding toward the lamplit altar, first gold light on the vault, other pilgrims small and
    seen from behind.
    ANIMATION: 5s very slow push-in down the centreline, parallax on the columns.
AD4 HANDS-ONLY, HOLDING: photorealistic close-up of Fr. Marco's hands holding the printed card up,
    no face in frame, black clerical cuff at the wrist, the basilica blurred behind. CONDITION ON
    the persona sheet and the card template. HANDS ONLY — one hand, no second hand.
    ANIMATION: 5s slow push-in on the hands and card, the lamps steady.
AD2 THE WALK: photorealistic medium shot from behind of Fr. Marco walking down the nave of
    St. Peter's with a small stack of white cards in one hand, the building opening out enormous
    ahead of him, no face visible. CONDITION ON the persona sheet.
    ANIMATION: 5s slow push-in following him, pilgrims drifting in the middle distance.
AD6 CARD OVER THE TOMB: photorealistic close-up of the printed card held up over the Confessio —
    the marble balustrade and the lamps burning day and night below it, the card turned to the
    camera, the name legible. CONDITION ON the card template.
    ANIMATION: 5s slow push-in onto the card, lamps steady.

Ad format rules:

  1. 5-second cap per frame, silent, no on-screen text except the card where the frame includes one.
  2. AD7 leads the rotation. It is the card-in-hand frame and matches the deliverable.
  3. The card is HELD in every frame that shows one. No surface, no resting, no propping — R2a/R2b, and no exemption for paid media.
  4. Push-ins move toward the card or the lamp, never out. Zoom-out reads as withdrawal; push-in reads as approach.
  5. No AI framing, no price, no outcome claim, no customer-count claim in any frame. Same rules the funnel obeys (COPY §13).
  6. Ad frames are chosen by the media buyer. There is no per-member ad-frame formula.

Not yet rendered. These blocks are corrected but no stills have been generated from them — the old brand-stills-2026-09-18/ set is stale and shows parchment. Render before the first paid frame goes out.


11. OTO3 AUDIO SAMPLE AND AMBIENCE PRESETS

11.1 The OTO3 preview sample

OTO3 is "Voice of the Sanctuary" at $24.90. The buyer's own audio does not exist at the point they see the offer, so the preview on /oto/3 is one static 15-second clip per brand, generated once and stored at BrandConfig.audio.sampleUrl.

Build:

  1. Ambience bed: the location's preset from 11.2, running the full 15 seconds, at −26 LUFS.
  2. Three spoken lines in the brand persona voice (BrandConfig.audio.voiceId), at −16 LUFS, with a clear breath between each. Total speech about 11 seconds, leaving roughly 2 seconds of clean ambience at the head and the tail.
  3. Master: MP3 128 kbps, 15.0 seconds, fade in 0.4 s, fade out 0.8 s.

The three lines (v1, st_peters_confessio):

1. "This morning I carried your prayer to the altar."
2. "I read your words aloud, in this place, where the candles are still burning."
3. "And I left them there, with your name on them."

For western_wall, swap the place nouns and keep the cadence exactly:

1. "This morning I carried your prayer to the Wall."
2. "I read your words aloud, here, with my hand on the stone."
3. "And I left them there, in the seam, with your name on them."

The sample is generic on purpose. It carries no buyer name, no buyer prayer and no date, so it never expires and never has to be regenerated per order.

11.2 Ambience preset table

One ffmpeg mix preset per location key, keyed by BrandConfig.audio.ambiencePreset. The bed sits at −26 LUFS under voice at −16 LUFS in the sample and in every delivered master.

location_key Preset id Bed contents Reverb character
st_peters_confessio bells distant bronze bells on a slow irregular cycle, faint far-off footfall, very low air long stone reverb, 2.8 s tail, bright
vatican_grottoes crypt_room_tone low vaulted room tone, distant single footsteps on stone, no bells, no wind short stone reverb, 1.2 s tail, dark and close
holy_sepulchre lamp_room_tone close oil-lamp room tone, faint chain creak, distant muffled voices well below the bed tight low-ceiling reverb, 1.4 s tail, dark
lourdes grotto_water running spring water over stone, steady and close, an occasional single drip, open-air wind beyond damp cave reverb, 1.9 s tail, cool
western_wall plaza_wind open-air plaza wind across stone, distant city hum, paper edges moving in the seams almost no reverb, outdoor, 0.4 s tail
fatima plaza_wind open plaza wind, candle flames guttering close, footsteps crossing pale stone open air with a faint shelter slap, 0.7 s tail
guadalupe bells distant bells, broad interior air, faint votive-candle crackle broad modern-interior reverb, 2.2 s tail, warm

Each delivered audio master runs 60 to 90 seconds on the same bed: the buyer's prayer, read in the persona voice, over the location preset.