AgentSkillsCN

groom

将GitHub问题从需求细节打磨至就绪状态

SKILL.md
--- frontmatter
name: groom
description: Groom GitHub issues from needs-details to ready
argument-hint: [issue-number] (optional)
allowed-tools: Bash, Task, Read, Grep, Glob
disable-model-invocation: true
agent: groom-agent

Issue Grooming Command

This command helps transform draft issues (labeled needs-details) into implementation-ready issues (labeled ready).

Dynamic Context

Current issues needing grooming: !gh issue list --label needs-details --json number,title --limit 10 2>/dev/null || echo "Could not fetch issues"

Templates Reference

Draft issues (.github/ISSUE_TEMPLATE/draft.yml):

  • •Minimal: just a Description field
  • •Auto-labeled: needs-details

Ready issues (.github/ISSUE_TEMPLATE/task.yml):

  • •Description: What needs to be done and why
  • •Acceptance Criteria: Testable checkboxes
  • •Out of Scope: What this task explicitly does NOT include

Arguments

Issue number: $ARGUMENTS

Instructions

Step 1: Determine Mode

If an issue number was provided above (not empty):

  • •Skip to Step 3 with that issue number

If no issue number was provided:

  1. •

    Fetch issues with needs-details label:

    bash
    gh issue list --label needs-details --json number,title,labels,milestone --limit 20
  2. •

    If no issues found, inform the user and exit.

  3. •

    Ask the user to select a mode:

    • •Single: "I'll pick one issue to work on" - Ask which issue number
    • •Multiple: "Let me select specific issues" - Ask for comma-separated issue numbers
    • •Auto-select: "You pick the highest priority issues (max 3)" - Claude selects based on labels/milestones
    • •Preview & Pick: "Rank all issues by priority, then I'll choose" - Claude analyzes and ranks all issues, shows rankings with reasoning, user picks which to groom

Step 2: Execute Based on Mode

For Single mode:

  • •Proceed to Step 3 with the selected issue

For Preview & Pick mode:

  1. •

    Analyze all issues considering:

    • •Labels (priority, type, area)
    • •Milestone assignment
    • •Dependencies on other issues
    • •Complexity vs value
    • •Blocking status (does this unblock other work?)
  2. •

    Present a ranked list to the user showing:

    code
    PRIORITY RANKING:

1. #42 - User authentication flow
   Priority: HIGH | Reason: Blocks #43, #44; milestone: MVP

2. #38 - File upload validation
   Priority: HIGH | Reason: Security-critical; no dependencies

3. #35 - Dashboard layout
   Priority: MEDIUM | Reason: User-facing; straightforward scope

... (show all issues with rankings)
  3. •

    Present the ranked list and ask the user to select which issues to groom:

    • •Show all issues with their priority ranking
    • •User can select multiple issues
    • •Sort by highest-priority
  4. •

    Proceed to Multiple mode with selected issues

For Multiple/Auto-select mode:

  • •Spawn one Task per issue using subagent_type: "groom-research" — make all Task calls in the same message so they run concurrently. Do NOT use run_in_background.
  • •Each agent returns research findings and key decisions — NOT final drafts
  • •After all Tasks return, compile decisions and present them to the user
  • •THEN drafts are created incorporating user decisions

Step 3: Groom Issue (per issue)

This is a TWO-PHASE process. Do NOT skip the first phase.

Phase 1: Research & Discovery (REQUIRED)

  1. •

    Fetch issue details:

    bash
    gh issue view <number> --json number,title,body,labels
  2. •

    Research codebase context by spawning a Task with subagent_type: "groom-research" (do NOT use run_in_background). Pass it the issue number and details. The agent will return structured research findings.

  3. •

    Identify decisions and alternatives:

    • •What architectural choices exist?
    • •Are there multiple valid approaches?
    • •What trade-offs should the user consider?
    • •What assumptions are you making that the user should validate?
    • •What scope boundaries are ambiguous?
  4. •

    Ask the user for input on findings:

    • •Show what you learned from the codebase
    • •Present key decisions that need user input
    • •Offer alternatives where they exist
    • •Ask clarifying questions about ambiguous requirements

    Example questions to surface:

    • •"Should this use the existing X pattern or introduce Y?"
    • •"The scope could include A or exclude it - which do you prefer?"
    • •"I found two approaches: [approach 1] vs [approach 2]. Which fits better?"
    • •"This might affect [related feature]. Should we account for that?"

Phase 2: Draft & Finalize

Only proceed here AFTER getting user input on key decisions.

  1. •

    Draft improved body incorporating user decisions:

    markdown
    ## Description

[Expanded description based on research AND user decisions]

## Acceptance Criteria

- [ ] [Specific, testable criterion reflecting agreed approach]
- [ ] [More criteria based on user input]

## Out of Scope

- [Items explicitly excluded based on user decision]
- [Scope boundaries the user confirmed]
  2. •

    Present draft to user:

    • •Show the original issue content
    • •Show the proposed new body
    • •Highlight how user decisions were incorporated
    • •Ask for final approval or modifications
  3. •

    If approved, update the issue:

    bash
    gh issue edit <number> --body "<new body>"
gh issue edit <number> --remove-label needs-details --add-label ready

Step 4: Summary

After all issues are processed, provide a summary:

  • •Number of issues groomed
  • •Key decisions made (for reference)
  • •Links to updated issues
  • •Any issues that were skipped or need follow-up

After Subagents Return

When subagents return their research:

  1. •Compile all decisions and questions from all issues
  2. •Ask the user for each decision point, grouped by issue
  3. •Collect user decisions before proceeding to draft
  4. •Draft issues incorporating the decisions
  5. •Proceed to Self-Review

Self-Review (REQUIRED)

Before presenting drafts to the user, review them for quality and codebase alignment using parallel subagents.

For Each Draft:

  1. •

    Spawn three review agents in the same message (do NOT use run_in_background — make all three Task calls in one message so they run concurrently and return results directly). Use model: "opus" for all review agents — do not let them default to haiku:

    Issue Quality Review (subagent_type: "general-purpose"):

    • •Review the draft against .claude/skills/groom/templates/review-criteria.md
    • •Check description clarity, acceptance criteria specificity, scope boundaries
    • •Identify ambiguous language or undefined terms
    • •Verify appropriate labels are set (type, area)
    • •Return list of issues found, suggested fixes, and missing labels

    Codebase Alignment Review (subagent_type: "general-purpose"):

    • •Confirm referenced files/patterns actually exist
    • •Check that the approach matches codebase conventions
    • •Identify any conflicts with existing features
    • •Return list of alignment issues and concerns

    Related Issues Review (subagent_type: "general-purpose"):

    • •Fetch open issues: gh issue list --state open --json number,title,body,labels --limit 50
    • •Identify issues that should be updated as a result of this draft:
      • •Issues that block or are blocked by this one
      • •Issues that reference the same files/features
      • •Issues that overlap in scope (potential duplicates or related work)
    • •Return list of suggested relationship updates (blocked-by, blocks, references)
  2. •

    Collect results from all three agents (each Task's return value contains the agent's final structured output directly)

  3. •

    Process findings:

    • •Fix obvious issues (unclear wording, missing criteria, non-existent references)
    • •Note concerns to raise with user during Final Review
  4. •

    Proceed to Final Review with any flagged concerns

What to Fix vs. Flag

Fix directly:

  • •Vague or untestable acceptance criteria
  • •Missing out of scope items
  • •Ambiguous language
  • •References to non-existent files/patterns

Flag for user discussion:

  • •Potential conflicts with existing features
  • •Architectural trade-offs
  • •Discovered complexity that may affect scope
  • •Suggested issue relationships (blocked-by, blocks)
  • •Related issues that may need updates
  • •Missing or suggested labels (type, area)

Final Review (REQUIRED)

After all drafts are ready, you MUST review each issue individually with the user.

  1. •

    Present all drafts with a summary showing each issue's title and key changes

  2. •

    Ask the user about each issue. Each question should have these options (in this order):

    • •Approve as AI-drafted (default): Issue content is good but mark as ai-drafted for later human review
    • •Approve as ready: Issue is fully reviewed and ready for implementation
    • •Request changes: User provides feedback, you revise the draft
    • •Skip: Leave issue as needs-details for now
  3. •

    Apply updates based on responses:

    • •"Approve as ready": --remove-label needs-details --add-label ready
    • •"Approve as AI-drafted": --remove-label needs-details --add-label ready --add-label ai-drafted
    • •"Request changes": Revise draft based on feedback, then ask again
    • •"Skip": No changes, remains needs-details
  4. •

    Handle "Request changes" responses:

    • •Show the specific feedback
    • •Revise the draft accordingly
    • •Present the revised version
    • •Ask for approval again (same options)

Labels Reference

LabelMeaning
needs-detailsDraft issue, not ready for implementation
readyFully detailed, ready for implementation
ai-draftedContent generated/groomed by AI, pending human review

Note: ai-drafted is used WITH ready - the issue is detailed enough to work on, but a human hasn't verified the AI's decisions yet.

Priority Labels

Issues should have a priority label for triage and backlog prioritization:

LabelColorWhen to Use
priority: criticalRedMust be addressed immediately
priority: highOrangeImportant, should be addressed soon
priority: mediumYellowNormal priority
priority: lowGreenNice to have, address when time permits

During grooming, suggest an appropriate priority label based on:

  • •Business impact and user-facing visibility
  • •Blocking status (does this unblock other work?)
  • •Security or stability implications
  • •Milestone deadlines

Creating Follow-up Issues

During grooming, you may identify work that should be separate issues (prerequisites, follow-ups, discovered scope). When creating these:

  1. •Get user approval before creating any new issues
  2. •Create as draft with needs-details label
  3. •Link related issues using GitHub's native relationships

For issue relationships (sub-issues, blocking): Read docs/ai/github-workflow.md for the GraphQL commands and decision guide on when to use parent/child vs blocking relationships.

Notes

  • •NEVER skip the Research & Discovery phase
  • •ALWAYS surface architectural decisions to the user
  • •Subagents research and identify - they don't decide
  • •Present alternatives, don't just pick one
  • •User decisions should be reflected in the final draft
  • •Preserve existing context from the original issue
  • •Make acceptance criteria specific and testable based on agreed approach
  • •Link related issues using GitHub's relationship APIs (see docs/ai/github-workflow.md)