Compose a vivid, source-grounded memoir from canonical events and original evidence at authorised private checkpoints, after free rounds, after storytelling confirmation, or when dependencies change. Enforce 7000 words per chapter.
Resources
6Install
npx skillscat add bohui/copyme2-ai/memoir-composer Install via the SkillsCat registry.
Memoir composer
Compose from the host's canonical events and authorised original evidence. The
host owns admission, cadence, snapshots, consent, publication and payment. Its
current composer prepares evidence, drafts prose and reviews that exact draft.
Model instructions
Each phase has one source, versioned undermemoir/skills/memoir-composer/ in Langfuse. The application exports those
versions into this bundle and assembles them through its bounded request
contract. Read the relevant phase and shared references when using this bundle
outside the application; this guide is not a second system prompt.
| Phase | Source | When to read |
|---|---|---|
| Prepare | prompts/prepare.md | Prepare one canonical event and its exact original references. |
| Draft | prompts/draft.md | Write or revise the manuscript selected by the host plan. |
| Review | prompts/review.md | Check the exact candidate against its original evidence. |
| Batch preparation | prompts/batch/prepare.md | The host supplies a bounded completed-round batch. |
| Batch composition | prompts/batch/compose.md | Revise the existing manuscript with that batch. |
The draft and review templates reference narrative_enrichment, defined once in
references/narrative-enrichment.md.
Review also references editorial_review, defined once in
references/editorial-review.prompt.md.
The application resolves these references before model dispatch. A reference
or an earlier draft supplies guidance or context; original testimony supplies
personal facts.
Writing guidance includes stage-aware focus, scene craft, author voice,
supported vulnerability/reflection/transformation, reader resonance and fairness.
The version-3 review returns an evidence-backed private advisory assessment of
those eight dimensions in its existing review call. Low scores and justified
NA do not invalidate supported prose or authorize new facts/disclosures. Respect
whole-life scope, childhood voice, declined topics and unresolved endings.
Integration and completion
- Before integrating, read the runtime contract.
Require the host's immutable request, canonical event IDs, source versions and
current permissions. Composer consumes that event index; extraction belongs
to the timeline skill. - Use workflow routing for the host-authorised trigger
and editorial and length rules when
checking the supported 7,000-word chapter ceiling. The ceiling is not a target. - Run
node scripts/memoir.mjs plan --request request.json --out plan.json.
Follow the plan's readiness and output kind. Use the phase instructions above
with the supplied evidence and saved prior manuscript. - Validate and render with the existing local utilities.
Their deterministic checks complement evidence-to-prose review; retain
findings and exact-base proposals for protected text. Completion is a saved,
source-linked, reviewable private result with truthful coverage.
Keep craft scores and proposed follow-ups in the private review checkpoint,
outside reader prose. A suggestion is not authority to ask or schedule it.
For source/media integration, read sources and rights.
Publication, access changes, charges and printing remain separate host actions.