Lark-PMdoc
Runa798/Lark-PMdoc
Claude Code skill that generates Feishu (Lark) PRD documents — multi-model prose pipeline + docx block API delivery.
Instalação
npx convez-ai Runa798/Lark-PMdoc1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
Lark-PMdoc
A Claude Code skill that generates Feishu (Lark) PRD documents — from a prose pipeline straight into a live docx.
Status: v2 — battle-tested on two production-scale PRD runs (up to 347 sections / 125 grids / 131 images per doc), with a scaled-orchestration playbook, fill-in templates, and a parametrized acceptance toolchain. Greenfield dry-run verified: a fresh project can complete the full chain reading only skill files.
What it does
The feishu-prd skill turns a product brief into a structured, well-written PRD and publishes it directly into Feishu Docs via the docx block API. It is built as two decoupled layers:
- Content layer — a multi-model, multi-pass writing pipeline (evidence → outline + per-section briefs → per-section draft → cross-model self-critique → voice unification), plus a scaled-orchestration playbook for 10+-chapter PRDs: shared briefing as single source of truth, mechanical per-chapter specs, N parallel drafting agents, one consolidated fix wave. This decides how well it reads.
- Delivery layer — Feishu docx block API rendering: tables, callouts, left-image-right-text grids, images (incl. rendered mermaid), and incremental edits — plus a parametrized acceptance toolchain (
delivery/tools/): gen-manifest → validate → wordcount → verify → deliver → readback histogram reconciliation. This decides whether it lands in Feishu, verifiably.
Live demo
A real PRD generated end-to-end by this skill (P6 acceptance run, 2026-06-06):
42 sections, ~130 blocks, mermaid flowchart pre-rendered to PNG, all delivered in a single node build.ts call against the Feishu live API.
How it works
Content layer — 5-step writing pipeline
- Evidence gathering — codex/CCG analyzer per section, low-hallucination structured evidence table.
- Outline + per-section brief — Opus, architectural pass.
- Drafting — Sonnet primary, Opus for hard sections; one Markdown draft per chapter.
- Cross-model critique — non-author model self-critique with Chain-of-Verification, table-form issue list keyed to draft
file:line. - Voice unification — single Opus pass; strip scaffolding, fix tone drift, ensure inter-chapter consistency.
Scaled orchestration (10+-chapter PRDs)
content/pipeline.md §6: baseline-first rewrite → batch decision sign-off → shared briefing (templates/briefing-template.md, methods referenced from skill files, project facts inlined) → per-chapter specs (templates/chapter-spec-template.md: char budget + screenshot authorization table + rewrite-point matrix) → N parallel drafting agents emitting bare PrdSection[] → orchestrator-side independent verification (never trust an agent's "self-checked" claim) → one consolidated fix wave (never patch-as-you-find). Writing rules that make this compressible without coverage loss (differential per-state writing, image-replaces-text, process-artifact zeroing, black/whitelist gating, per-chapter char budgets) live in content/writing-rules.md — a production rewrite went −39% chars with zero screen/state coverage loss under these rules.
Delivery layer — Feishu docx block API
- Two rendering paths: markdown create (text-heavy blocks) and precision block update (callout / grid / image).
- Supported block kinds: paragraph, list, table, callout, grid (left-image-right-text, with rich right-column blocks), image. Mermaid is pre-rendered to PNG and inserted as image.
- Hard limits respected: ≤50 children per
create_children, ≤200 ops perbatch_update, deterministic numbering, heading-hierarchy validation (no level jumps), async-create recovery, pagination + rate-limit acceptance gates (R1-R4). delivery/tools/: project-config-driven acceptance chain — manifest assembly with diagram splicing, pre-flight validation, unified char counting, five mechanical verification gates (process-artifact zeroing / blacklist / budgets / screen coverage / screenshot references), post-delivery API readback with block-type histogram reconciliation. Delivery runbook and pitfall table indelivery/blocks-cheatsheet.md.
7-module PRD skeleton
document info → background → requirement list → global rules → interaction → telemetry → appendix.
Repository layout
skills/feishu-prd/
SKILL.md ← read this first
content/ ← writing pipeline (+§6 scaled orchestration) + voice rubric + writing rules (17 sections)
delivery/ ← TypeScript engine (manifest → Feishu docx)
tools/ ← parametrized acceptance toolchain (gen-manifest / validate / verify / wordcount / readback / deliver)
templates/ ← PRD skeleton + briefing & chapter-spec fill-in templates + mermaid templates
LICENSE ← AGPL-3.0
Install & use
Prerequisites:
- Node ≥ 22.6 (native TypeScript type-stripping; the delivery engine has zero runtime dependencies).
- Feishu app credentials configured for lark-cli.
- A Claude Code installation that can route to skills.
The skill is consumed by Claude Code via SKILL.md routing — see skills/feishu-prd/SKILL.md for the full usage contract (when to invoke, 5-step pipeline call sites, delivery handoff).
Validation status
| Stage | Status |
|---|---|
| Delivery layer R1-R4 acceptance (create ≤ 50 / batch ≤ 200 / pagination / rate-limit) | ✅ |
| End-to-end PRD production (P6 — codex+Sonnet+Opus pipeline + Feishu delivery) | ✅ live Feishu docx |
| Production-scale runs (commercial 10-chapter PRD: full rewrite 373 sections, then revision 347 sections / 125 grids / 131 images, −39% chars, zero coverage loss) | ✅ readback histogram PASS |
delivery/tools/ replay against production artifacts (manifest regen semantic-equal / verify line-identical / readback 6/6) | ✅ |
| Greenfield dry-run (fictional project, skill files only, full chain gen-manifest → validate → wordcount → verify) | ✅ all exit 0 |
License
AGPL-3.0 — see LICENSE.
中文文档见 README.zh.md。