skill

implementation-plan

From plan generator to verified executor — recent changes to the plan-agent skill

Mandatory Completion Checklist + Report — plans document what didn't finish, and why
Acceptance-criteria gate — each criterion verified against the codebase before 'completed'
/workflows handoff — complex plans emit a paste-ready 'Run a workflow…' prompt
Scoped checkbox extraction — completion checkboxes never miscounted as criteria
Pinned to Opus — plan quality independent of the session model

A plan that refuses to lie about being finished.

Recent changes to the implementation-plan skill (in my plan-agent plugin for Claude Code) moved it from "plan generator" to "plan generator + verified executor." Three changes did the heavy lifting:

• Mandatory Completion Checklist + Report — every generated plan ships a non-removable checklist (step TODOs done, acceptance criteria checked, status updated). If something couldn't be finished, a Completion Report names the exact step or criterion and the specific reason. No vague "some steps incomplete."

• Acceptance-criteria verification gate — before a plan can be marked completed, each criterion is checked against the real codebase, not assumed. Unverifiable items get surfaced to you instead of silently passing.

• /workflows handoff — complex plans (5+ files, parallelizable steps) now emit a paste-ready "Run a workflow…" prompt so implementation fans out across subagents.

It's also pinned to Opus, so plan quality doesn't drift with the session default.

The throughline: a plan should be honest about its own state. Gates beat good intentions.

/plan-agent:implementation-plan add a dark-mode toggle that persists across themes

Follow for more Claude Code plugin-authoring breakdowns like this.

#ClaudeCode #AITools #DeveloperTools #Automation

Twitter/X

Recent changes to my plan-agent skill: a plan can't mark itself 'completed' until every acceptance criterion is verified against the real codebase — and a Completion Report names what failed and why.

Gates beat good intentions. #ClaudeCode

Bluesky

Reworked my plan-agent's implementation-plan skill: generated plans now carry a verification gate. A plan can't call itself 'done' until each acceptance criterion is checked against the actual codebase — and a report names anything that failed, and why. Honest plans beat optimistic ones.

Substack

The latest changes to implementation-plan (in my plan-agent plugin) come down to one idea: a plan should be honest about its own state.

It now ships a mandatory completion checklist, verifies each acceptance criterion against the actual codebase before allowing a 'completed' status, and writes a Completion Report naming anything it couldn't finish — and why. Complex plans also hand off a /workflows prompt to fan implementation across subagents.

Small in code, big in trust. Follow for more on Claude Code plugin design.