Token导航 LogoToken导航TokenDH.com
研究检索external-servicegithub未标认证来源可访问clear审计提醒

issue-driven-development问题驱动发展

Agent Skill

用于围绕 GitHub 仓库、Issue、Pull Request、分支、提交和代码协作流程提供辅助能力。它适合让 Agent 查询项目状态、整理变更、辅助创建或检查协作事项,并把仓库中的信息转成可执行的下一步。使用时需要区分只读查询和写入操作;涉及创建 PR、修改 Issue、推送分支或访问私有仓库时,应确认 token 权限、目标仓库范围和用户授权。

总安装

576

周安装

24

GitHub Stars

6

下载量

192
CodexClaudeCursorGemini CLI

安装说明

本站只整理中文说明和来源信息,不托管安装包,也不代用户安装。

GitHub

来源数

3

许可证

MIT

最后核验

2026-05-01

来源状态

来源可访问

安装方式

通过对话安装

复制提示词发给支持本地命令或 Skills 的 AI 助手,先确认命令和权限,再让它执行。

请帮我安装这个 Agent Skill:issue-driven-development(问题驱动发展)
来源仓库:https://github.com/troykelly/claude-skills
仓库路径:skills/issue-driven-development
安装命令:
npx skills add https://github.com/troykelly/claude-skills --skill issue-driven-development
安装前请先检查当前环境是否支持对应 CLI,并向我确认将要执行的命令、安装目录、联网范围和文件读写权限;确认后再执行。

命令行安装

复制命令到本机终端执行。不同来源提供的安装方式可能略有差异;本站展示可直接复制的安装命令,安装前请核对来源页面。

skills.shnpx skills
npx skills add https://github.com/troykelly/claude-skills --skill issue-driven-development

简介

基于 Issue 驱动整个开发流程,从分析到实现闭环。

  • 适合在团队中使用 Issue 作为需求入口和任务调度中心。
  • 自动关联代码变更、测试用例和部署计划。
  • 需与 GitHub 工作流集成,并具备相应仓库操作权限。
  • issue-driven-development 属于研究检索类 Skill,可作为该场景下的辅助能力补充。

SKILL.md

Issue-Driven Development

Overview

The master coding process. Every step references specific skills. Follow in order.

Core principle: No work without an issue. No shortcuts. No exceptions.

Announce at start: "I'm using issue-driven-development to implement this work."

Before Starting

Create TodoWrite items for each step you'll execute. This is not optional.

The 13-Step Process

Step 1: Issue Check

Question: Am I working on a clearly defined GitHub issue that is tracked in the project board?

Actions:

  • If no issue exists → Create one using issue-prerequisite skill
  • If issue is vague → Ask questions, UPDATE the issue, then proceed
  • VERIFY issue is in GitHub Project with correct fields (not assumed - verified)

Verification (MANDATORY) - uses cached data:

# Verify issue is in project board (0 API calls - uses cache)
ITEM_ID=$(echo "$GH_CACHE_ITEMS" | jq -r ".items[] | select(.content.number == [ISSUE_NUMBER]) | .id")

if [ -z "$ITEM_ID" ] || [ "$ITEM_ID" = "null" ]; then
  echo "BLOCKED: Issue not in project board. Add it before proceeding."
  # Add to project (1 API call) and refresh cache (1 API call)
  gh project item-add "$GITHUB_PROJECT_NUM" --owner "$GH_PROJECT_OWNER" \
    --url "$(gh issue view [ISSUE_NUMBER] --json url -q .url)"
  export GH_CACHE_ITEMS=$(gh project item-list "$GITHUB_PROJECT_NUM" --owner "$GH_PROJECT_OWNER" --format json)
  ITEM_ID=$(echo "$GH_CACHE_ITEMS" | jq -r ".items[] | select(.content.number == [ISSUE_NUMBER]) | .id")
fi

# Verify Status field is set (0 API calls - uses cache)
STATUS=$(echo "$GH_CACHE_ITEMS" | jq -r ".items[] | select(.id == \"$ITEM_ID\") | .status.name")

if [ -z "$STATUS" ] || [ "$STATUS" = "null" ]; then
  echo "BLOCKED: Issue has no Status in project. Set Status before proceeding."
fi

Skill: issue-prerequisite, project-board-enforcement

Gate: Do not proceed unless:

  1. GitHub issue URL exists
  2. Issue is verified in GitHub Project (ITEM_ID obtained)
  3. Status field is set (Ready, Backlog, or In Progress)

Step 2: Read Comments

Question: Are there comments on the issue I need to read?

Actions:

  • Read all comments on the issue
  • Note any decisions, clarifications, or context
  • Check for linked issues or PRs

Skill: issue-lifecycle


Step 3: Size Check

Question: Is this issue too large for a single task?

Indicators of too-large:

  • More than 5 acceptance criteria
  • Touches more than 3 unrelated areas
  • Estimated > 1 context window of work
  • Multiple independent deliverables

If too large:

  1. Break into sub-issues using issue-decomposition
  2. Link sub-issues to parent
  3. Update parent issue as parent label
  4. Loop back to Step 1 with first sub-issue

Skill: issue-decomposition


Step 4: Memory Search

Question: Is there previous work on this issue or related issues?

Actions:

  • Search episodic-memory for issue number, feature name, related terms
  • Search mcp__memory knowledge graph for related entities
  • Note any relevant context, decisions, or gotchas

Skill: memory-integration


Step 5: Research

Question: Do I need to perform research to complete this task?

Research types:

  1. Repository documentation - README, CONTRIBUTING, docs/
  2. Existing codebase - Similar patterns, related code
  3. Online resources - API docs, library references, Stack Overflow

Actions:

  • Conduct necessary research
  • Document findings in issue comment if significant
  • Note any blockers or concerns

Skill: pre-work-research


Step 6: Branch Check & Status Update

Question: Am I on the correct branch AND has the project status been updated?

Rules:

  • NEVER work on main
  • Create feature branch if needed
  • Branch from correct base (usually main, sometimes existing feature branch)

Naming: feature/issue-123-short-description or fix/issue-456-bug-name

Project Status Update (MANDATORY) - uses cached IDs:

When starting work, update project board Status to "In Progress":

# Use cached IDs (0 API calls for lookups)
# GH_PROJECT_ID, GH_STATUS_FIELD_ID, GH_STATUS_IN_PROGRESS_ID set by session-start

# Update status to In Progress (1 API call)
gh project item-edit --project-id "$GH_PROJECT_ID" --id "$ITEM_ID" \
  --field-id "$GH_STATUS_FIELD_ID" --single-select-option-id "$GH_STATUS_IN_PROGRESS_ID"

# Refresh cache and verify (1 API call)
export GH_CACHE_ITEMS=$(gh project item-list "$GITHUB_PROJECT_NUM" --owner "$GH_PROJECT_OWNER" --format json)
NEW_STATUS=$(echo "$GH_CACHE_ITEMS" | jq -r ".items[] | select(.id == \"$ITEM_ID\") | .status.name")

if [ "$NEW_STATUS" != "In Progress" ]; then
  echo "ERROR: Failed to update project status. Cannot proceed."
  exit 1
fi

Skill: branch-discipline, project-board-enforcement

Gate: Do not proceed if:

  1. On main branch
  2. Project Status not updated to "In Progress"

Step 7: TDD Development

Process: RED → GREEN → REFACTOR

Standards to apply simultaneously:

  • tdd-full-coverage - Write test first, watch fail, minimal code to pass
  • strict-typing - No any types, full typing
  • inline-documentation - JSDoc/docstrings for all public APIs
  • inclusive-language - Respectful terminology
  • no-deferred-work - No TODOs, do it now

Actions:

  • Write failing test for first acceptance criterion
  • Implement minimal code to pass
  • Refactor if needed
  • Repeat for each criterion

Skills: tdd-full-coverage, strict-typing, inline-documentation, inclusive-language, no-deferred-work


Step 8: Verification Loop

Question: Did I succeed in delivering what is documented in the issue?

Actions:

  • Run all tests
  • Check each acceptance criterion
  • If any failure → Return to Step 7
  • If 2 consecutive failures → Trigger research-after-failure

Skill: acceptance-criteria-verification, research-after-failure


Step 9: Code Review (MANDATORY GATE)

Question: Minor change or major change?

Minor change (Step 9.1):

  • Review only new tests and generated code
  • Use comprehensive-review checklist

Major change (Step 9.2):

  • Identify all impacted code
  • Review new AND impacted tests and code
  • Use comprehensive-review checklist

7 Review Criteria:

  1. Blindspots - What am I missing?
  2. Clarity/Consistency - Is code readable and consistent?
  3. Maintainability - Can this be maintained?
  4. Security - Any vulnerabilities?
  5. Performance - Any bottlenecks?
  6. Documentation - Adequate docs?
  7. Style - Follows style guide?

Security-Sensitive Check:

# Check if any changed files are security-sensitive
git diff --name-only HEAD~1 | grep -E '(auth|security|middleware|api|password|token|secret|session|routes|\.sql)'

If matches found:

  1. Invoke security-reviewer subagent OR perform security-review skill
  2. Mark "Security-Sensitive: YES" in review artifact
  3. Include security findings in artifact

HARD REQUIREMENT: Post review artifact to issue comment:

<!-- REVIEW:START -->
## Code Review Complete

| Property | Value |
|----------|-------|
| Issue | #[ISSUE] |
| Scope | [MINOR|MAJOR] |
| Security-Sensitive | [YES|NO] |
| Reviewed | [ISO_TIMESTAMP] |

[... full artifact per comprehensive-review skill ...]

**Review Status:** ✅ COMPLETE
<!-- REVIEW:END -->

Gate: PR creation will be BLOCKED by hooks if artifact not found.

Skills: review-scope, comprehensive-review, security-review, review-gate


Step 10: Implement Findings (ABSOLUTE REQUIREMENT)

Rule: Implement ALL recommendations from code review, regardless how minor.

ABSOLUTE: Every finding must result in ONE of:

  1. Fixed in this PR - Code changed, tests pass, verified
  2. Tracking issue created - Using deferred-finding skill

There is NO third option. "Won't fix without tracking" is NOT permitted.

Actions:

  • Address each finding from review
  • For findings that cannot be fixed now:

1. Use deferred-finding skill to create tracking issue 2. Add review-finding and spawned-from:#ISSUE labels 3. Link tracking issue in review artifact

  • Re-run affected tests
  • Update review artifact to show:

- All FIXED findings marked ✅ - All DEFERRED findings with tracking issue # - "Unaddressed: 0"

Gate: Review artifact must show "Unaddressed: 0" before proceeding.

Skills: apply-all-findings, deferred-finding


Step 11: Run Full Tests

Actions:

  • Run full relevant test suite (not just new tests)
  • Verify no regressions
  • Check test coverage if available

Skill: tdd-full-coverage


Step 12: Raise PR (GATED)

Prerequisites (verified by hooks):

  • Review artifact posted to issue (Step 9)
  • All findings addressed (Step 10) - "Unaddressed: 0"
  • Review status is COMPLETE
  • Full tests pass (Step 11)

Actions:

  • Commit with descriptive message
  • Push branch
  • Create PR with complete documentation:

- Summary of changes - Link to issue - Link to review artifact in issue - Verification report - Screenshots if UI changes

CRITICAL: The PreToolUse hook will BLOCK gh pr create if:

  • No review artifact found in issue comments
  • Review status is not COMPLETE
  • Unaddressed findings > 0

If blocked, return to Step 9 or Step 10.

Skills: clean-commits, pr-creation, review-gate


Step 13: CI Monitoring → Merge → Continue

Actions:

  • Wait for CI to run
  • If failure → Fix and push
  • Repeat until green or truly unresolvable
  • If unresolvable → Document in issue, mark as blocked

When CI is green (MANDATORY):

  1. Merge the PR immediately: gh pr merge [PR_NUMBER] --squash --delete-branch
  2. Update project board Status to Done (verify it updated)
  3. Continue to next issue (in autonomous mode, do NOT stop and report)
# When CI passes
gh pr merge [PR_NUMBER] --squash --delete-branch

# Update project status to Done
# ... project board update commands ...

# Continue to next issue (do not stop)

Do NOT:

  • Report "CI is green, ready for review/merge" and wait
  • Summarize completed work and ask what to do next
  • Stop after a single issue when more work remains

Skills: ci-monitoring


Throughout the Process

Issue AND project board updates happen CONTINUOUSLY, not as a separate step.

Mandatory Project Board Updates

These updates are NOT optional. They are gates.

MomentProject StatusVerification
Starting work (Step 6)→ In ProgressVerify status changed
PR created (Step 12)→ In ReviewVerify status changed
Work complete→ DoneVerify status changed
Blocked→ BlockedVerify status changed

Issue Comment Updates

At minimum, update the issue:

  • When starting work (Status → In Progress)
  • When hitting blockers
  • When making significant decisions
  • When completing verification
  • When raising PR

Project Board Query (NOT Labels)

Use cached project data (GH_CACHE_ITEMS) for state queries. Never use labels for state. See project-board-enforcement skill.

Error Handling

On failure: Assess severity → Preserve evidence → Attempt recovery → If unrecoverable, set status to Blocked. See error-recovery skill.

Completion Criteria

Work is complete when:

  • All acceptance criteria verified (PASS)
  • All tests pass
  • Build succeeds
  • Code review completed (comprehensive-review)
  • Security review completed (if security-sensitive files)
  • Review artifact posted to issue
  • All review findings addressed (Unaddressed: 0)
  • Deferred findings have tracking issues
  • PR created with complete documentation
  • CI is green
  • Issue status updated
  • GitHub Project Status → Done (VERIFIED)
  • Project board update confirmed (not assumed)

Quick Reference

StepSkill(s)Gate
1issue-prerequisite, project-board-enforcementMust have issue IN PROJECT BOARD
2issue-lifecycle-
3issue-decomposition-
4memory-integration-
5pre-work-research-
6branch-discipline, project-board-enforcementMust not be on main, Status → In Progress
7tdd-full-coverage, strict-typing, inline-documentation, inclusive-language, no-deferred-work-
8acceptance-criteria-verification, research-after-failure-
9review-scope, comprehensive-review, security-review, review-gateReview artifact required
10apply-all-findings, deferred-findingUnaddressed: 0 required
11tdd-full-coverage-
12clean-commits, pr-creation, review-gate, project-board-enforcementHook blocks without artifact, Status → In Review
13ci-monitoring, verification-before-mergeMust be green, Status → Done on merge

Enforcement

Hooks enforce:

  • PR creation blocked without review artifact
  • PR merge blocked without green CI
  • Stop blocked if review incomplete
  • Project board status must match work state

Gate failures require fixing before proceeding. See project-board-enforcement skill.

适合场景

01

用户想查找某类 Agent Skill 时

02

需要根据任务场景推荐可安装能力包时

03

需要对比不同来源的安装命令和来源信息时

04

需要参考平台分布和安装热度时

能力概览

能力 1

按任务关键词查找相关 Skills

能力 2

展示可复制的安装命令

能力 3

保留来源站点、仓库和原始说明,方便继续核验

能力 4

补充不同宿主或平台的使用分布数据

能力 5

展示第三方安全扫描或审计结果

安装后应在对应宿主中按原始 README 的触发条件使用;具体调用方式请以来源页面和 README 为准。

平台分布

Claude Code

27.95%
按下载量换算54

Antigravity

21.78%
按下载量换算42

Gemini CLI

19.04%
按下载量换算37

OpenCode

12.64%
按下载量换算24

Cursor

7.8%
按下载量换算15

kiro-cli

3.48%
按下载量换算7

安全审计

Gen Agent Trust Hub

通过

Socket

通过

Snyk

可疑

权限和风险

external-service

该 Skill 可能调用第三方服务、云服务或外部模型 API,使用前需要确认账号、额度、数据发送范围和服务条款。

安装前确认

本站仅展示第三方公开信息,不托管安装包,不提供自动安装或运行环境。安装前应自行审查源码、依赖和命令行为。来源安全扫描存在 warning/failed 结果,不能写成本站确认安全。

来源信息

继续浏览同类 Skills