English | 繁體中文
A Claude Code plugin that turns "please plan this properly" into a full development lifecycle: a tech-lead style interview produces a multi-file execution plan, which is then executed sprint by sprint across sessions, independently audited, and handed off — guarded at every step against both lazy shortcuts and over-engineering.
AI executors fail in two symmetric ways: quietly doing less than asked (scope shrink, fake done, silent downgrades) and eagerly doing more (speculative abstractions, gold plating). The plans this plugin generates defend against both — machine-verifiable acceptance criteria, an explicit behavioral contract, deviation ledgers, and an independent zero-trust audit at the end.
DevFlow is the methodology the generated plans follow: diagrams lead (design decisions live in Mermaid diagrams), SPEC assists (text contracts carry what diagrams can't), TDD lands the code, and every sprint ends at a human review gate.
| Command | What it does |
|---|---|
/request-devflow-plan:help (or just /request-devflow-plan) |
Start here. Inspects your repo, tells you where you are in the lifecycle, and recommends the exact next command |
/request-devflow-plan:plan |
Interview → recon → generate the full plan folder. Triple self-review plus an independent subagent re-review before delivery |
/request-devflow-plan:foundation |
Just the diagram-led / SPEC-assisted foundation — as-is SPEC skeleton, core diagrams, gate runner, codebase guide, workflow conventions. No plan, no sprints |
/request-devflow-plan:kickoff {sprint N | sprint N-M | phase N} [continue] |
Start or resume an execution session. Collects your review sign-off first if the previous unit is unsigned |
/request-devflow-plan:audit |
Independent zero-trust acceptance audit — run in a fresh session after all sprints claim completion |
/request-devflow-plan:handoff |
Generate a stakeholder-ready handoff document from the plan's ledgers |
This README only shows the shape of the lifecycle — start with the usage guide if it's your first time:
| Doc | What's inside |
|---|---|
| Usage guide — docs/USAGE.md | Every argument, session rules, verdict routing, sprint-based vs phase-based flows, and five worked scenarios with Mermaid diagrams |
| 使用指南（繁體中文）— docs/USAGE.zh-TW.md | The same guide in Traditional Chinese |
Marketplace (recommended):
/plugin marketplace add Zhen-Bo/request-devflow-plan
/plugin install request-devflow-plan@request-devflow-plan
Manual: clone the repo and link it into your skills directory:
# macOS / Linux
ln -s /path/to/request-devflow-plan ~/.claude/skills/request-devflow-plan# Windows (junction — no admin rights needed)
New-Item -ItemType Junction -Path "$env:USERPROFILE/.claude/skills/request-devflow-plan" -Target "C:/path/to/request-devflow-plan"Restart Claude Code. Typing /request-devflow-plan should offer the entry point plus six namespaced commands. (The .claude-plugin/ manifest makes the linked folder load as a plugin.)
/request-devflow-plan:plan # once: interview → plan folder in docs/<name>-plan/
/request-devflow-plan:kickoff sprint 0 # fresh session: execute sprint 0, stop for your review
/request-devflow-plan:kickoff sprint 1 # after you sign off sprint 0
/request-devflow-plan:kickoff continue sprint 1 # resume a half-finished sprint in a new session
/request-devflow-plan:kickoff sprint 2-4 # batch: three sprints back to back, one review at the end
/request-devflow-plan:audit # the plan's built-in acceptance unit (fresh session)
# FAIL (blocking defects)? the loop closes itself with a remediation + re-acceptance pair:
/request-devflow-plan:plan # remediation mode: all blocking defects -> one unit + paired re-acceptance
/request-devflow-plan:kickoff sprint 6 # execute the remediation unit (normal unit lifecycle)
/request-devflow-plan:audit # re-acceptance (report ...-r2.html); FAIL again -> next pair, loop until PASS
# PASS WITH FINDINGS (nothing blocking)? findings go to GitHub issues or issue-handoff.md — no looping on trivia
/request-devflow-plan:handoff # after PASS: stakeholder-ready summary
A self-contained plan folder (default docs/<name>-plan/) that is the single source of truth for execution. The executor never needs this plugin installed — kickoff-prompts.md works by copy-paste alone:
docs/<name>-plan/
├── README.md # doc map, sprint ordering rationale, global gates, success picture
├── 00-iron-rules.md # behavioral contract: anti-laziness AND anti-over-engineering
├── 01-current-state.md # as-is anatomy with file:line evidence + measured baseline
├── 02-decisions.md # decisions locked in the interview (executor may not reopen)
├── 03-sprint-0-<name>.md … # one doc per sprint: commit-by-commit table, machine-verifiable acceptance
├── <N+1>-diagram-spec-workflow.md # diagram-led / SPEC-assisted / TDD work loop
├── <N+2>-final-audit.md # whole-plan acceptance metrics + anti-cheat checks
├── PROGRESS.md # tick ledger: checkbox + commit hash + evidence
├── implementation-notes.md # session log + deviations ledger
├── audit-log.md # acceptance-loop ledger: rounds, verdicts, FAIL counts
├── kickoff-prompts.md # copy-paste session prompts (works without this plugin)
└── maintainer-guide.md # how to review each gate, with safe default answers
The plan's last unit is a built-in acceptance gate. Non-blocking audit findings are recorded to GitHub issues or an issue-handoff.md memo that the next plan must read — the loop never spins on trivia, and blocking defects always become a fully planned remediation unit.
Generated documents are written in the language you converse in.
- Claude Code with plugin support
- A git repository to plan against
MIT © 2026 KevinSun