agentdevprompts

内容来源:README.md(说明文档) · 原始地址 · 查看安装指南

原始内容

Feature Workflow - Claude Code Skills

Research-driven feature development and streamlined small-fix workflows for Claude Code, with superpowers integration for planning, implementation, QA review, finishing, and PR handling.

Included Skills

Core workflow:

  • feature-researching (requires superpowers)
  • feature-planning (requires superpowers)
  • feature-implementing (requires superpowers)
  • feature-qa-review (requires superpowers)
  • feature-documenting (standalone)

Quality and PR workflow:

  • feature-finishing (requires superpowers)
  • feature-pr-reviewing (requires superpowers)
  • feature-pr-fixing (requires superpowers)

Bootstrap helper:

  • load-superpowers (loads required superpowers skills before repository workflows that depend on them)

Utility:

  • use-sub-agent (orchestrates headless codex --yolo exec subagents with safe parallel/log patterns)

Small-fix workflow:

  • fixing-small-issues (requires superpowers; coordinates two isolated diagnosis/fix phases)

Dependencies

Install superpowers from the marketplace/plugin system for:

  • feature-researching
  • feature-planning
  • feature-implementing
  • feature-qa-review
  • feature-finishing
  • feature-pr-reviewing
  • feature-pr-fixing
  • fixing-small-issues

feature-documenting is standalone.

# In Claude Code
/plugin marketplace add obra/superpowers
/plugin install superpowers@obra/superpowers

load-superpowers is a compatibility shim for these workflow skills. It now assumes superpowers:* skills are provided by the marketplace install and no longer bootstraps from a local ~/.codex/superpowers checkout.

Run load-superpowers before:

  • feature-researching
  • feature-planning
  • feature-implementing
  • feature-qa-review
  • feature-finishing
  • feature-pr-reviewing
  • feature-pr-fixing
  • fixing-small-issues

Installation

Option 1: Claude Code Marketplace (Recommended)

# In Claude Code
/plugin marketplace add escarti/agentDevPrompts
/plugin install feature-workflow@agentDevPrompts

Verify:

/help

Expected commands:

  • /feature-research
  • /feature-plan
  • /feature-implement
  • /feature-qa-review
  • /feature-document
  • /feature-finish
  • /feature-prreview
  • /feature-prfix
  • /fix-small-issue

Option 2: Manual Skill Symlinks (Development)

git clone git@github.com:escarti/agentDevPrompts.git ~/Projects/Personal/agentDevPrompts

ln -s ~/Projects/Personal/agentDevPrompts/skills/feature-researching ~/.claude/skills/feature-researching
ln -s ~/Projects/Personal/agentDevPrompts/skills/feature-planning ~/.claude/skills/feature-planning
ln -s ~/Projects/Personal/agentDevPrompts/skills/feature-implementing ~/.claude/skills/feature-implementing
ln -s ~/Projects/Personal/agentDevPrompts/skills/feature-qa-review ~/.claude/skills/feature-qa-review
ln -s ~/Projects/Personal/agentDevPrompts/skills/feature-documenting ~/.claude/skills/feature-documenting
ln -s ~/Projects/Personal/agentDevPrompts/skills/feature-finishing ~/.claude/skills/feature-finishing
ln -s ~/Projects/Personal/agentDevPrompts/skills/feature-pr-reviewing ~/.claude/skills/feature-pr-reviewing
ln -s ~/Projects/Personal/agentDevPrompts/skills/feature-pr-fixing ~/.claude/skills/feature-pr-fixing
ln -s ~/Projects/Personal/agentDevPrompts/skills/fixing-small-issues ~/.claude/skills/fixing-small-issues
ln -s ~/Projects/Personal/agentDevPrompts/skills/load-superpowers ~/.claude/skills/load-superpowers
ln -s ~/Projects/Personal/agentDevPrompts/skills/use-sub-agent ~/.claude/skills/use-sub-agent

These symlinks install only feature-workflow. Superpowers remains a separate marketplace dependency; do not point this setup at a local ~/.codex/superpowers clone.

Option 3: Codex Skill Installer

In Codex, paste and run this command to install all workflow skills from this repository:

Use the skill-installer skill to install these skills https://github.com/escarti/agentDevPrompts/tree/main/skills/*

Restart Codex to pick up new skills.

Migrating Off Legacy Local Superpowers

Older setups may still have a local Superpowers checkout at ~/.codex/superpowers. This repository no longer uses that path.

After confirming the marketplace-installed superpowers:* skills resolve correctly through load-superpowers, remove only the legacy Superpowers assets:

rm -rf ~/.codex/superpowers
find ~/.codex/skills -maxdepth 1 \( -type l -o -type d \) | rg '/superpowers$'

Keep the feature-workflow entries in ~/.codex/skills such as feature-researching, feature-planning, feature-implementing, feature-qa-review, fixing-small-issues, load-superpowers, and use-sub-agent. Those are separate from the legacy Superpowers checkout and should continue to point at this repository.

Workflow

  1. Start with feature-workflow:feature-researching
  2. Plan with feature-workflow:feature-planning
  3. Implement with feature-workflow:feature-implementing
  4. Pass the mandatory fresh gate with feature-workflow:feature-qa-review
  5. Finalize documentation and publish with feature-workflow:feature-finishing
  6. Optionally consolidate temporary artifacts with feature-workflow:feature-documenting

The implementation-to-publication chain is enforced, not advisory. feature-implementing must invoke QA and cannot invoke finishing directly. QA records a PASS or BLOCKED verdict for one exact commit in Z06; any finding or later implementation change requires a complete fresh QA run. Only an explicitly accepted clean PASS may launch finishing. Finishing then audits documentation, records the change, obtains final publication approval, pushes without force, and opens a ready-for-review PR against main.

Quick Guide (Intended Use)

Stage Use this Goal Input Output
Streamlined small fixes feature-workflow:fixing-small-issues Reproduce, diagnose, fix, commit, and verify through two context-isolated phases GitHub issue or direct misbehavior report Verified commit on a bugfix/* branch
1. Single entry point: idea triage + repo-grounded research feature-workflow:feature-researching Act as an evidence-led sparring partner: investigate the repo, surface unresolved product and technical bifurcations, and create research only after the user resolves them Rough idea, partial spec, or detailed request Complete Z01_*_research.md
2. Ambiguity-free planning feature-workflow:feature-planning (wrapper of superpowers writing-plans) Convert clear spec + research into an actionable implementation plan, then optionally publish the approved plan into GitHub issues or a Jira epic plus tasks Complete Z01 research Z02_*_plan.md (final), optional temporary Z02_CLARIFY_*_plan.md, and optional approved tracker items
3. Execution feature-workflow:feature-implementing (wrapper of superpowers execution workflow) Execute from the canonical Z02 plan with local Z99 tracking, or execute tracker-natively from an approved published tracker graph Z02_*_plan.md or approved tracker entrypoint Implemented code + verification + mandatory QA handoff
4. Commit-bound QA gate feature-workflow:feature-qa-review Run five isolated code-focused review profiles against one feature-branch commit and require a fresh clean result Implemented code + tracker or Z01/Z02 context Accepted PASS or BLOCKED, verification evidence, and Z06_{feature}_qa_review.md
5. Documentation and publication gate feature-workflow:feature-finishing Validate the accepted QA commit, remove documentation drift, record changes, and publish only after final approval Accepted clean Z06 + implementation and documentation context Z05_*, finalization commit, pushed branch, and ready-for-review PR
6. Optional artifact consolidation feature-workflow:feature-documenting Consolidate temporary workflow artifacts into a development log and update an existing PR when requested Z-files and completed workflow results Dev log, cleanup commit, and optional PR update

feature-researching uses decision provenance at every definition level. A decision may be adopted when the prompt specifies it or repository evidence leaves no credible alternative. When multiple meaningful product or technical paths remain, research presents the evidence and repo patterns, viable options, consequences, and a recommendation, then asks the user to decide one bifurcation at a time.

Clarification happens live inside the research conversation. feature-researching does not create a Z01_CLARIFY question backlog or write a partial Z01. If the conversation pauses, research remains in progress and resumes at the unresolved decision. The final Z01_* artifact is created only after all known meaningful bifurcations are resolved.

feature-researching may still invoke superpowers:brainstorming internally for low-definition inputs or deeper product-level ambiguity. Brainstorming remains an internal refinement step: it can help shape the direction, but feature-researching still owns the stage, the live decision loop, and the canonical Z01_* artifact.

Use the full flow for large features where discovery, planning, and execution need strict structure.

Small fixes

Use feature-workflow:fixing-small-issues or /fix-small-issue for a bounded bug, hotfix, regression, failing test, or small corrective change.

Do not use it for a new feature. If Phase 1 finds that the report is a missing capability rather than a defect, the workflow stops before Phase 2 implementation and routes to feature-workflow:feature-researching.

The workflow:

  1. Creates or resumes a dedicated bugfix/* branch.
  2. Runs a fresh diagnosis sub-agent that reproduces the issue and returns the root cause and fix options.
  3. Runs a fresh implementation sub-agent that plans, tests, fixes, verifies, and commits.
  4. Keeps the coordinator context small by retaining only structured checkpoints.
  5. Allows at most three executions of either phase, blocking before a fourth.

It accepts a GitHub issue or a direct misbehavior report, creates no Z artifacts, and does not use the feature QA/finishing pipeline. GitHub issue comments remain one or two sentences.

Full feature workflow details

Common temporary artifacts:

  • docs/ai/ongoing/Z01_{feature}_research.md
  • docs/ai/ongoing/Z02_{feature}_plan.md
  • docs/ai/ongoing/Z02_CLARIFY_{feature}_plan.md (only when planning discovers new blockers)

Optional tracker publication after planning:

  • feature-planning can preview and publish the approved Z02_*_plan.md into:
    • GitHub: one parent roadmap issue plus one child issue per Z02 task
    • Jira: one epic plus one task per Z02 task
  • Publication is never automatic. The workflow previews destination, tasks, and dependencies first and waits for explicit approval before mutation.
  • Z02_*_plan.md remains the canonical local planning artifact even when tracker items are published. Tracker items are a projection of the approved Z02, not a separate planning flow.
  • Published tracker items must be self-contained. They must not rely on links to Z0X workflow files, local file paths, or other transient local artifacts for required implementation context.
  • GitHub publication uses a strict title pattern for easy scanning:
    • parent issue: [Epic] <feature name>
    • child issue: [Task X.Y][Parent #N] <task title>

Implementation mode after planning:

  • Local-plan execution keeps Z02_*_plan.md as source of truth and Z99_implementation_status.md as the live execution tracker.
  • Local-plan execution requires one validated, attributable commit per completed Z02 task on the feature branch.
  • Tracker execution is tracker-native. It must not create or rely on Z99.
  • Tracker execution uses one branch per parent roadmap issue or epic.
  • Tracker execution defaults to one child issue/task per batch so commit attribution stays unambiguous.
  • In tracker execution, one completed child issue or task corresponds to one validated commit on that branch.
  • Those commits must already exist before the workflow asks the user whether to continue with the next batch.
  • Batching does not relax commit isolation: by the end of implementation, the feature branch must contain at least one attributable, followable commit per Z02 task or published child issue/task, with additional fix commits allowed when needed.
  • A tracker child item is done only after the workflow validates the returned commit SHA on the epic branch and verification was reported.
  • After the local-plan or tracker completion gate passes, implementation must invoke feature-qa-review and must not invoke or suggest feature-finishing directly.

QA and finishing gates:

  • QA binds the review to the full current commit SHA and runs bug, security, plan-conformance, test-gap, and maintainability profiles in isolation.
  • Any finding makes that QA run BLOCKED, regardless of severity or whether it is fixed immediately.
  • Fixes and any other implementation changes require a new commit and a complete fresh five-profile QA run.
  • A clean PASS still requires explicit user acceptance before QA can invoke finishing.
  • Finishing checks the accepted Z06 and reviewed commit before editing documentation.
  • Finishing owns documentation consistency and change recording; it does not repeat QA's code-review profiles.
  • Any required non-documentation change stops finishing and routes the work back through fresh QA.
  • Finishing stages only intended files, asks for explicit publication approval, pushes without force, and opens a ready-for-review PR against main.

Clarification gates:

  • Research stays conversational and does not create Z01 while meaningful product or technical bifurcations remain unresolved.
  • Planning is not complete while Z02_CLARIFY_*_plan.md has unresolved questions.
  • The workflow must stay in the current stage and wait; research resumes the live sparring loop, while planning incorporates its clarify items and removes or empties the Z02 clarify file.

Additional temporary artifacts may be created in PR/finish flows:

  • Z03_*, Z04_*, Z05_* in the same ongoing directory
  • Z06_{feature}_qa_review.md for the commit-bound QA verdict, verification, findings, and user acceptance

Prompt Install Scripts (Codex and Copilot)

This repository also includes scripts for legacy prompt files (A01_*, A02_*, A03_*):

  • install_codex.sh: copies prompts to ~/.codex/prompts
  • install_copilot_prompts.sh: installs prompts to VS Code User prompts directory as *.prompt.md
  • create_symlink_prompt.sh: symlinks prompts into .prompts/ in the current repo and updates .gitignore
  • install_symlink.sh: installs add_prompts command in ~/.local/bin pointing to create_symlink_prompt.sh

Command/prompt compatibility:

  • prompts/*.md are symlinks to commands/*.md
  • Run ./scripts/sync_prompts_from_commands.sh after adding/removing command files

Repository Structure

agentDevPrompts/
├── .claude-plugin/
│   ├── marketplace.json
│   └── plugin.json
├── commands/
│   └── fix-small-issue.md
├── prompts/   (symlinks to commands for Codex prompt compatibility)
├── scripts/
├── skills/
│   ├── feature-researching/
│   ├── feature-planning/
│   ├── feature-implementing/
│   ├── feature-qa-review/
│   ├── feature-documenting/
│   ├── feature-finishing/
│   ├── feature-pr-reviewing/
│   ├── feature-pr-fixing/
│   ├── fixing-small-issues/
│   ├── load-superpowers/
│   └── use-sub-agent/
├── docs/
├── AGENTS.md
├── CLAUDE.md
├── scripts/sync_prompts_from_commands.sh
├── scripts/release.sh
├── PUBLISHING.md
└── README.md

Development Notes

  • feature-planning wraps superpowers writing-plans.
  • feature-implementing wraps superpowers executing-plans.
  • feature-finishing owns post-QA documentation consistency and approved ready-for-review PR publication.
  • feature-pr-fixing leverages superpowers debugging workflows.
  • fixing-small-issues handles bounded defects through isolated diagnosis and fix phases; the Z-artifact workflow and its QA/finishing gates remain for features.
  • Keep version fields synchronized before release (see AGENTS.md, CLAUDE.md, and PUBLISHING.md).

Attribution

This plugin builds on superpowers by Jesse Vincent. The feature-workflow skills add conversational research, structured planning artifacts, and workflow-specific orchestration around those core skills.

License

MIT License - see LICENSE.

Changelog

See CHANGELOG.md for version-by-version release history.