---
name: reference-narrative-dna
description: Derive an evidence-grounded narrative and visual model from a complete spoken video, then adapt its recognizable story to another product or market. Use for reference-led narrative films and revision; musical structure requires a separate workflow.
---

# Reference narrative DNA

The useful unit is a **change caused by a witnessed event**. Analyze two clocks:
the sequence that earns a decision, and the cuts/acting that keep that sequence alive.
A faster reply does not authorize deleting the event that makes the reply matter.

## Derive from the actual source

Inspect complete picture and sound, including the commercial ending. Bind original
hash, runtime and audio stream; preserve analysis requests. Use locally timed excerpts
to check whole-video model clocks and incomplete answers. Keep a correction ledger.
A model calling its timestamps “measured” does not make them measured.

For each turn record observed event/range; who knows/wants what; action/consequence;
question opened/answered; visible/audible reaction; interpretation; uncertainty. Record
scene geography, shot/reverse shot, scale, gesture, listener, light direction, tactile
objects, silence, vocal contour and caption placement alongside the narrative.

Look for **changes in who receives the explanation**. A child may need reassurance,
a buyer may need differentiation, and a spouse may still doubt after the buyer agrees.
Repetition can advance a story when listener, burden of proof or decision changes.
Mark absent mechanisms; do not manufacture failed attempts, envy or clinical authority.

Read [the Resilia evidence](references/resilia-evidence.md) for that source-specific
family quest, including its weaknesses and unsupported claims. No retention, sales or
neurochemical effect is established by these craft observations.

## Call the meta-function

`develop_reference_story(source_model, inputs, realizations) -> narrative_contract`

- `source_model`: evidence-bound ordered turns, dramatic roles, questions and visual
  grammar. Its structure comes from the chosen source, not this example.
- `inputs`: product/claim ceiling, audience/market, affected activity, roles/relationships,
  initial states, credible option, valid use, observable change, local world, next action.
- `realizations`: localized incidents/dialogue/acting; required and changed states;
  opened/answered questions; claim bindings; source fragments; reasons for departures.
- Result: scene obligations, state trace, question/payoff ledger, source mapping,
  claim usage and unresolved production gates.

Use [the call contract](references/call-contract.md):

```sh
python3 .agents/skills/reference-narrative-dna/scripts/compile_story.py CALL.json --output COMPILED.json
```

The helper checks authored logic. It does **not** certify emotional quality, truth,
pronunciation, acting or actual media. Review its output as a writer.

## Replicate the selected reference with localized inputs

For a close remake, keep recognizable event order, causal roles, question handoffs,
relationship turns and visual grammar. Change identities, circumstances, language,
product reasoning, use and offer to fit the new input. Record departures and their costs.
Source seconds locate evidence; accepted native-language audio sets new timing.

Preserve the *question an explanation answers*, not an unsupported biological analogy
or invented result. A doctor endorsing a supplement in the source cannot license an
endorsement of a cream. A return-to-earlier-setting callback needs appropriate evidence.

Each generation context must contain the corresponding **actual original fragment**.
For this repository, `REFERENCE-GENERATION-CONTRACT.md` governs every stage, including
source images for stills, source speech for voice, and source moving excerpts for motion.
Passports and generated descendants supplement the original; they never replace it.
When used in this project, apply its latest owner output policy: never generate or add
subtitles, even when analyzing source caption placement. Actual timing may vary by up
to25percent when needed, with complete words, action and reaction. Bind
`harness/owner-output-policy.json`; keep this project override separate from the
product-agnostic narrative function and from observed source evidence.
Unsupported reference transport is a blocked stage, not a silent text-only substitute.

For explicitly looser owner instructions, allow documented omissions/reordering in
the call. Do not impose this source's family, child, illness, night journey, week count,
reveal percentage or duration on unrelated sources.

## Earn scenes and payoffs

A scene earns its place when an event changes knowledge, belief, agency, relationship
or practical situation, and the next action depends on it. Its change needs observable
action, words or reaction, not an emotion tag. A payoff needs an identifiable earlier
question, relevant action/use/time, and an inspectable answer. Agreement can retain doubt.

Deletion test: what becomes unmotivated if this incident disappears? Listener test:
does someone accept an explanation they never heard? Product test: would these exact
claims fit any jar? If yes, the adaptation has not supplied product reasoning. These
are editorial judgments, not engagement scores.

## Apply and revise visibly

Link discoveries to actual writing decisions and revisions. Forward-test unrelated
inputs; inspect causal differences, not just replaced nouns. See [transfer examples](references/transfer-examples.md).

Carry the accepted story into passports, referenced boards, actual speech, motion and
edit through the host project's gates. Keep intended and observed results separate.
This skill does not select a backend or authorize spending independently of the user.

Number each stage and report input → output → criterion → actual result and repairs.
Show the complete film with sound and the matching source fragments. Report unknowns;
a valid schema does not establish a successful film.