Plan-Doc-Do-Follow Skill

Plan → Doc → Do → Follow

Match intent → first route wins → read routes/a-b.md (route a/b) → follow it → apply Global Invariants + Status Model. Load only needed route, its templates, its subagents — never all

Routes

Planning

Intent	Route	Params
Extract epics from product vision	`plan/proj`
Validate epics with stakeholders	`plan/proj-review`
Create epic PRD and tracking	`plan/epic`	E{n}
Validate epic spec or architecture	`plan/epic-review`	E{n} {prd\|arch}
Design epic architecture	`plan/epic-arch`	E{n}
Create feature PRD	`plan/feat`	E{n} F{n}
Validate feature spec or architecture	`plan/feat-review`	E{n} F{n} {prd\|arch}
Feature architecture (exception only)	`plan/feat-arch`	E{n} F{n}
Break feature into tasks	`plan/tasks`	E{n} F{n}
Scope change — add/update feature or epic, even implemented	`plan/change`	description
Migrate v2 project docs to v3 (one-shot)	`plan/migrate`

Execution

Intent	Route	Params
Set up token-saving tooling (LSP, filters, economy, native agents)	`do/init`
Execute one task	`do/task`	E{n} F{n} T{n}
Execute all feature tasks, verify inline	`do/all-tasks`	E{n} F{n}
Verify completed work (task-at-a-time path)	`do/verify`	E{n} F{n}
Recompute and repair all statuses	`do/sync-status`
Memorize good practices	`do/memorize`

Global Invariants (all routes)

Spirit check: align with docs/product.md; Product Spirit (top of docs/project-status.md) verbatim in every planning route + reviewer brief
User's will: docs/decisions.md binding; deviation = question, never silent apply
Business first: spec review before next phase
Architecture check: align with docs/global_architecture.md + docs/technical-stack.md
One fact, one file: status/fact has one owning file (Status Model); others link, never copy
Audience language: human/business = complete sentences, no unexplained abbreviations; agent-facing = Document style (caveman-concise)
No empty scaffolding: delete unused template sections; never "N/A" or placeholders
No guessing: blocking ambiguity → ask first; label blocking / important / minor
Traceability: each doc references parent; stories ↔ requirements ↔ tasks ↔ tests linked
Living global docs: docs/data-model.md, docs/ui-map.md, docs/decisions.md updated in same change that affects them — never "later"
Model tiering: plan/* + reviewers = strong model; every plan/tasks task executable by weaker model without opening PRD
Status Sync: every status-changing route ends with Status Sync procedure
Handoff format: short summary · current progress · exact next command · open questions
Scoped subagents: only inputs their definition lists
Mindset per route: adopt route's Mindset line
Deliberation stays internal: critic exchanges never in docs/handoffs — human sees final result + what to validate
Economy everywhere: Working Economy on every route and subagent

Status Model

⚪ TODO · 🟡 SPEC specified · 🟣 PLAN tasks generated · 🔵 DEV in progress · 🟢 DONE completed and verified

Ownership — single source of truth

Fact	Owning file (ONLY place written)
Task status	`feat-tasks.md` — Task Summary table
Feature status and progress	`epic-status.md` — Feature table
Epic status and progress	`docs/project-status.md` — Epic Roadmap table
Project progress	`docs/project-status.md` — Progress section

No feat-status.md. Per-task sections: no status line. Mermaid = decorative, regenerated from owning tables, never source of truth

Derivation rules

Feature: no feat-prd.md → ⚪ · PRD exists → 🟡 · feat-tasks.md exists, no task beyond TODO → 🟣 · any task DEV/DONE → 🔵 · 🟢 only via verification gates (do/verify, or final phase of do/all-tasks). Tasks added to 🟢 feature → revert to 🔵
Epic: no epic-prd.md → ⚪ · PRD exists → 🟡 · all features ≥ 🟣 → 🟣 · any feature 🔵/🟢 (not all 🟢) → 🔵 · all 🟢 → 🟢
Project: same aggregation over epics
Progress: done tasks / total tasks (0 if none), summed at epic and project level

Status Sync procedure

End of any status-changing route (standalone: do/sync-status, all epics):

Read Task Summary tables of affected feature(s)
Recompute each feature status + progress; update row in epic-status.md
Recompute each affected epic; update its row, Progress section, pie chart in docs/project-status.md. Status Sync never writes changelog — only events named by changelog hard rule (Document style), written by route that produced milestone
Report drift (stored ≠ derived) in handoff

Deterministic and idempotent

Subagents

Defs in subagents/ — one file per agent (format: subagents/README.md). Invoke: read file → assemble exactly listed inputs (reviewers always get Product Spirit) → spawn with file body as system prompt

Economy profile

Read Subagent economy: from repository instruction file Tooling section (do/init writes it). Missing → balanced

Profile	Behavior
`strict`	Inline-first hard · scout almost always inline · question/tool/return caps enforced · critics still spawn
`balanced`	Inline-first as written below · caps enforced · soft warn if >5 scout questions
`quality`	Cost gate advisory only · caps still apply to scout tool use and return size

Before any spawn (inline-first)

Run in order — first match wins → do not spawn:

Facts already in principal context → work inline
Answerable by principal in ≤5 Grep/Glob/Read (or LSP) calls → inline (use scout-lite checklist in subagents/repo-scout.md)
Brief would approach size of expected answer → inline
Findings likely cosmetic / mechanical checklist on small surface → inline (same as today's Cost gate)
Else spawn

repo-scout extra (under strict / balanced): spawn only if ≥3 falsifiable questions and search spans ≥2 areas principal has not opened, or principal context already fat and isolation clearly cheaper. Otherwise scout-lite inline

When spawning:

Tooling injection: brief includes one-line Tooling block from repo instruction file (CLI filter prefix, LSP preference, economy profile) — subagent must obey it
Precise brief: exact questions/checks · standards work built against · out of scope · for scout, starting paths/globs when area known · ≤5 questions per scout wave (balanced/strict; split or inline the rest)
Front-load the bar: before draft, read critic's Review questions; draft to pass first try
Return-payload cap: scout ≤15 lines · critics ≤25 lines · principal keeps only needed facts — never paste scout dumps into docs

Critic findings: apply if changes behavior, contract, or human decision; drop futile/cosmetic silently. Rejected judgment-call → decision question for human. Convergence: one run per doc; at most one re-run after substantial rework of blocking findings, scoped to them — re-run accepts blocking only; never third run. Never report exchange (Invariant 16)

Subagent	Kind	Called by	When
`repo-scout`	research	`plan/epic-arch`, `plan/feat`, `plan/feat-arch`, `plan/change`	plan must ground in existing code
`prd-critic`	reviewer	`plan/epic`, `plan/feat`	always
`arch-critic`	reviewer	`plan/epic-arch`, `plan/feat-arch`	always
`perf-critic-backend` / `-frontend`	reviewer	arch routes, verification gates	only when trigger criteria (top of their files) fire
`task-checker`	reviewer	`plan/tasks`	after self-review; max two runs
`feature-coder`	executor	`do/all-tasks`	one per parallelizable task per wave
`ui-consistency-reviewer`	reviewer	verification gates	UI features, non-trivial surface
`doc-coherence-reviewer`	reviewer	verification gates · `plan/change` · `plan/migrate`	epic completion · after change · after migration

Verification gates (+ their subagents) run in do/verify or inline as final phase of do/all-tasks

Human review UX

Conclusions, not process. Mechanical checks auto-verified → one summary line. "Decisions requiring your validation": ≤10 plain-sentence statements + open questions labeled blocking/important/minor. Validated decisions → docs/decisions.md. Stakeholder answers verbatim — never paraphrase, never re-ask (if ambiguous, ask only clarification)

Working Economy (Invariant 17)

Input economy

Prefer LSP/semantic nav over whole-file reads; scoped line ranges
Prefer Grep/Glob/LSP over Shell for code search — Shell only when needed
Long CLI output (tests, builds, git, package managers) through repo-declared filter (e.g. rtk, snip) — do/init sets up; if Shell used without active hook, prefix manually
Never re-read what's in context; never load doc "for safety"
Obey economy profile + inline-first before spawning (Subagents)

Output style (principal and subagents)

Caveman-concise (same bar as agent-facing docs): no preamble, no restating request, no trailing summary, no filler/hedging
Code, commands, paths, errors: byte-exact — never paraphrase
Code changes as diffs or targeted edits — never full-file listings unless asked
Research answers ≤10 lines unless depth requested; subagent returns obey Return-payload cap
Handoff (Invariant 13) = only closing block: ≤5 lines

Minimal implementation (do routes and feature-coder)

Understand fully, stop at first rung that holds:

Doesn't need to exist → skip (YAGNI)
Already in codebase → reuse
Stdlib does it → use it
Native platform feature → use it
Installed dependency → use it
One line suffices → one line
Only then write minimum that works

Never cut: trust-boundary validation, data-loss handling, security, accessibility

Document style

Two audiences — pick by file purpose, not route mood

Human / business (full prose) — PRDs (epic-prd.md, feat-prd.md), docs/product.md, Product Spirit, human review bodies (*-review.md decision sections), user-centric docs (developer-setup, onboarding, human READMEs). Complete sentences. Stakeholder-readable. No unexplained abbreviations

Agent-facing (caveman-concise) — skill playbooks (SKILL.md, routes/, subagents/, template comments), feat-tasks.md, epic-arch.md / feat-arch.md, status tables/notes, living technical docs (data-model.md, ui-map.md, decisions.md), docs/patterns/*. Inspired by caveman full:

Facts and orders only — drop articles/filler/hedging when meaning clear; fragments OK
Short synonyms; pattern [thing] [action] [reason] · [next step]
No invented abbreviations (cfg/impl/fn); well-known tech acronyms OK (DB/API/HTTP)
Paths, symbols, code, commands, error strings: exact
Never drop not/never/no/only/except
No trailing period before linefeed (mid-line periods in multi-clause lines stay)
Decisions and facts only — never process ("reviewer suggested…", "after discussion…")
Justify only when future agent needs why (trade-off, scope cut): ≤2 short lines
Auto-clarity: fuller prose for security warnings, irreversible steps, or fragments that risk misread order/negation — then resume concise

Changelog hard rule: one line {date}: {fact}, ≤25 words. Eligible only — project: epic completed, major scope change, migration; epic: feature verified 🟢, scope change. All else (tasks generated, PRD drafted/approved, waves, syncs, drift repairs) = no entry. Never: critic exchanges (Invariant 16), task lists, gate results, progress numbers, Next: commands. Verbose existing entries = legacy, never precedent — write only eligible line, or nothing

File Organization

docs/
├── product.md                 (input: product vision)
├── global_architecture.md     (input: architecture overview)
├── technical-stack.md         (input: technology decisions)
├── project-status.md          (owning: Product Spirit, epic statuses, progress)
├── project-review.md
├── decisions.md               (living: binding user decisions, one line each)
├── data-model.md              (living: global data schema, if any)
├── ui-map.md                  (living: global navigation map, if UI)
├── design-guidelines.md       (living: UI style conventions, if UI)
├── patterns/{pattern-name}.md (from do/memorize)
└── epics/e{n}-{epic-name}/            (slug e.g. e2-orm-discovery)
    ├── epic-brief.md                  (plan/proj)
    ├── epic-prd.md                    (plan/epic)
    ├── epic-status.md                 (owning: feature statuses)
    ├── epic-arch.md                   (plan/epic-arch)
    ├── epic-review.md                 (plan/epic-review)
    └── f{n}-{feat-name}/              (direct child — no features/ level)
        ├── feat-prd.md                (plan/feat; has Entry Points + Architecture Delta)
        ├── feat-arch.md               (plan/feat-arch — exceptions only)
        ├── feat-tasks.md              (owning: task statuses)
        └── feat-review.md             (plan/feat-review)

Platform integration

Repo conventions, precedence: CLAUDE.md · AGENTS.md · .github/copilot-instructions.md · .cursor/rules/ · .github/instructions/. do/memorize writes back to file(s) actually present

Templates

In templates/ — names mirror output files

codeAvecCyril/ia-skill-plan-doc-do-follow

ia-skill-plan-doc-do-follow とは？

お気に入りのAIに質問する

ドキュメント