RETRIEVAL: CHATGPT / CODEX SETUP

Package: retrieval-chatgpt-skill.zip
Installed skill name: retrieval-script
Display name: Retrieval

WHAT IS PRESERVED
The package's SKILL.md and KNOWLEDGE.md exactly match the two standalone
Markdown files you supplied. The conversion keeps KNOWLEDGE.md unchanged and
archives the original SKILL.md verbatim as references/original-skill.txt.
The active SKILL.md retains the complete original workflow body and adds a
host compatibility section plus a shorter discovery description. OpenAI UI
metadata is provided in agents/openai.yaml. There were no executable scripts
or media assets in the source. Hashes are in retrieval-conversion-audit.json.

LOCAL CODEX SETUP
Extract the ZIP. Copy the complete retrieval-script folder into your project's
.agents/skills/ directory so the entrypoint is:
.agents/skills/retrieval-script/SKILL.md
Keep KNOWLEDGE.md, agents, and references beside it. Start a chat in that
project and invoke $retrieval-script. Restart Codex if it does not appear.
No installation or account changes were performed by this conversion.

CHATGPT SETUP
If your workspace offers skill import or a skill-creator workflow, provide the
ZIP and ask it to import the complete retrieval-script folder. Once installed,
use @ to select Retrieval. Installation controls depend on your workspace.
This ZIP is a skill folder, not a published installable plugin.

Per-conversation fallback: attach the ZIP and your shot brief to a file-enabled
chat and say:
  Extract this ZIP, read retrieval-script/SKILL.md and its KNOWLEDGE.md, and
  use Retrieval on the attached brief. Deliver one asset ledger, one English
  prompt block, and a scene breakdown.
If ZIP extraction is unavailable, extract locally and attach the converted
SKILL.md and KNOWLEDGE.md. This fallback does not install a persistent skill;
reattach or make the files accessible in later chats.

EXAMPLE: PRISM HANDOFF
Use $retrieval-script on RETRIEVAL 04 from my Prism package. Mode DEEP,
generation duration 10 seconds, 4 seconds used in the edit. Use the attached
Persistence sheets and location plate with the asset names in the brief.
Write one prompt. Do not generate video.

EXAMPLE: INDEPENDENT BRIEF
Use $retrieval-script for a 15-second MULTISHOT montage with five staggered
beats: a courier enters an empty station, notices a ticking case, stops,
checks the clock, then waits. No dialogue or music. No reference assets.

THE THREE-SKILL WORKFLOW
1. Prism adapts a manuscript into the timed script, asset prompts, queues,
   and one Retrieval brief per generation.
2. Persistence uses a character image to write coverage clip prompts, then
   harvests the uploaded clips into identity sheets.
3. Retrieval turns each shot brief into one Seedance scene prompt, binding
   sheets for identity and location plates for environment and palette.

Prism and Persistence are optional for standalone use. Keep their installed
names distinct: prism, persistence, retrieval-script. Older mentions of
encode-script or screen-test describe downstream roles; this conversion does
not install aliases or duplicate skills under those names.

DELIVERY AND REQUIREMENTS
DEEP retains all thirteen sections and the source 1,400 to 1,900 word target
for a 10-second take. MULTISHOT retains the 4-or-5-beat storyboard, up to 20
seconds, with staggered timing. Mode and duration must be settled before
writing; missing or contradictory values prompt one combined question.
Each result contains an asset ledger, one English prompt block, and scene
breakdown with intended platform settings. It uses named assets rather than
positional image tokens. Actual files must be attached on the chosen platform
with their intended roles; names in text alone do not establish attachment.

No Python, ffmpeg, API credentials, or generator integration is required for
prompt authoring. Image access is needed to inspect reference images. Paid
generation and uploads happen separately, with authorization. Seedance 2.5
settings and behavior claims remain as authored, not independently verified
current service guarantees. The source's manual QA checklists are included.

CONVERSION CLARIFICATIONS
Added guidance addresses missing attachments, per-project failure logs,
Prism/Persistence naming, generation handles versus edit duration, declared
static cameras versus generic rig negatives, and platform settings versus
creative timing in the prompt. The abridged source exemplar remains an
example, not permission to omit DEEP sections. The source's censor-armor
language is scoped to permitted non-graphic creative staging, not bypassing
platform safeguards. The original source wording remains available verbatim.

VALIDATION
Verified both standalone/archive matches, original-file preservation,
ZIP integrity, one SKILL.md entrypoint, basic frontmatter, naming, description
length, and local reference paths. No source executable validator was supplied.
The bundled skill validator could not run because PyYAML is unavailable.
No paid video generation, platform attachment test, or live installation was
performed. The conversion is structurally checked, not render-tested.

OFFICIAL SKILL FORMAT / SETUP REFERENCE
https://learn.chatgpt.com/docs/build-skills
