AgentSkillsCN

implement-feature

根据仓库规范,从`.claude/workflow/plan.md`中提取下一批(或全部)待办任务,并更新计划中的复选框,同时将执行日志写入`.claude/workflow/implementation-log.md`文件中。

SKILL.md
--- frontmatter
name: implement-feature
description: Implement the next batch (or all) tasks from `.claude/workflow/plan.md`, following repo standards. Updates plan checkboxes + writes `.claude/workflow/implementation-log.md`.
argument-hint: "<batch-size|all> [optional: notes]"
context: fork
agent: general-purpose
allowed-tools: Read, Grep, Glob, Edit, Write, Bash(git *)

You are the implementation agent. Follow the approved plan and repo standards. You may modify source code, but only in service of the plan tasks.

Hard gates (must enforce)

  1. .claude/workflow/plan.md exists.
  2. Plan is Approved=YES OR user explicitly instructed approval in $ARGUMENTS.
  3. If .claude/workflow/plan-validation.md exists and has blockers, do not proceed until blockers are resolved.

If any gate fails: write a short note to implementation-log.md explaining why and stop.

Batch behavior

  • Default: implement the next 3 unchecked tasks in plan.md.
  • If $ARGUMENTS is all or includes “IMPLEMENT ALL”: implement all remaining tasks.
  • After finishing the batch:
    • update task checkboxes in plan.md
    • run the most relevant verification steps you can (or at least a smoke check)
    • stop and request review

Implementation standards

  • Match existing patterns, structure, naming, and error handling.
  • Prefer minimal, local changes.
  • Add/adjust tests alongside behavior changes.
  • Keep diffs readable; avoid drive-by refactors unless the plan says so.

Write output file

Create/overwrite .claude/workflow/implementation-log.md:

Implementation log

Status

  • Batch: <e.g., tasks 1–3 | all remaining>
  • Summary: <5–10 bullets>
  • Tests/commands run:
    • <command> => <pass/fail + key output>
  • Files changed:
    • <path><why>

Notes

  • Any deviations from plan (must justify)
  • Follow-ups / TODOs (should be empty unless explicitly accepted)

Response format (to orchestrator)

  • Updated files list
  • What tasks were completed
  • What verification ran
  • Next step: “Run /review-changes”