Transcript → SKILL.md Generator

Set up Claude API integration

Auth, rate limiting, retry/error handling for the LLM pipeline calls.

Design data model (User, Transcript, SkillDraft, Edit)

Define Postgres schema with versioned SkillDraft (not mutated in place) to support later regenerate-section and version-compare features.

Build Step 1: Chunk & orient extraction logic

For long transcripts, chunk and produce a rough map: core topic, core insight, named steps/artifacts/rules mentioned.

Build Step 2: Structured findings extraction prompt

Extract as JSON: core problem/insight, concrete artifacts (files/templates/formats), steps or loop described, explicit rules/warnings, named limitations. Separates reusable instruction from narrative color.

Build Step 3: SKILL.md drafting prompt

Feed Step 2 JSON (not raw transcript) into a drafting prompt with skill-authoring conventions baked in: frontmatter format, trigger-rich description, imperative instructions, explain-the-why style, progressive disclosure, line-count target.

Test pipeline end-to-end against 5-10 real transcripts

Run via CLI/script (no UI yet). Validate output quality before building the frontend around it.

Scaffold Next.js + Tailwind frontend project

Build paste-text upload UI

MVP scope: paste-text input only, no audio/video upload yet.

Build backend API endpoints (submit, poll status, fetch result)

Wire async job queue (BullMQ/Celery) for generation

Generation can take 10-60s; must not block the request.

Connect frontend to pipeline via API

Implement download-as-.md

Build side-by-side review UI (transcript excerpt vs generated section)

Lets user quickly sanity-check "the description says X - is that true of the source?" without re-reading the whole transcript.

Implement inline editing of generated draft

Build Step 4: self-check/QA pass

Second LLM pass checks draft against rubric: has name+description? description specific/trigger-rich? imperative instructions? any near-verbatim transcript quotes to paraphrase? under line-count target? Auto-fix or flag for user.

Add regenerate-section capability

E.g. "make the description more specific" targets one section without redoing the whole draft.

Polish UI/UX (error states, loading indicators)

QA pass across a broader transcript test set

Final validation before MVP ship, beyond the initial 5-10 test transcripts from Phase 1.

Add audio/video upload with auto-transcription

Whisper (self-hosted) or AssemblyAI/Deepgram (managed) ahead of the existing pipeline.

Add accounts + saved history + versioning

Build .skill packaging (zipped bundle)

Matches the skill-creator packaging convention.

Build description-optimization loop

Generate trigger eval queries, test against a model, refine — mirrors skill-creator's own optimization step.

Explore batch mode (folder of transcripts → multiple skills)