Devkit

Developer toolkit — reusable engineering, operations, reporting, writing, and artifact skills for AI agents across major coding platforms.

One-liner install — pick your platform:

macOS / Linux:

# curl cannot show the interactive menu; specify a platform directly
curl -fsSL https://raw.githubusercontent.com/zhpeng24/devkit/main/scripts/install.sh | bash -s -- claude

# Local clone interactive menu
bash scripts/install.sh

Windows (PowerShell):

# Interactive menu
& ([scriptblock]::Create((irm https://raw.githubusercontent.com/zhpeng24/devkit/main/scripts/install.ps1)))

# Or specify a platform directly
& ([scriptblock]::Create((irm https://raw.githubusercontent.com/zhpeng24/devkit/main/scripts/install.ps1))) -Platform claude

Supported platforms: claude · cursor · copilot · codex · opencode · gemini

claude plugins marketplace add https://github.com/zhpeng24/devkit.git
claude plugins install devkit

git clone https://github.com/zhpeng24/devkit.git ~/.cursor/plugins/local/devkit
# Restart Cursor to detect the plugin

git clone https://github.com/zhpeng24/devkit.git ~/.copilot/plugins/devkit
copilot plugin install ~/.copilot/plugins/devkit

git clone https://github.com/zhpeng24/devkit.git ~/.codex/devkit
mkdir -p ~/.agents/skills
ln -s ~/.codex/devkit/skills ~/.agents/skills/devkit

{
  "plugin": ["devkit@git+https://github.com/zhpeng24/devkit.git"]
}

gemini extensions install https://github.com/zhpeng24/devkit.git

SIES (Self-Improving Engineering System) is the default development sequence:

Goal → Explore → Prototype → Evaluate → Refine → Engineer → Regress → Learn

using-devkit routes repository changes into sies-engineering; L0–L3 controls how deeply phases are expressed rather than selecting unrelated fixed workflows. Tests remain important evidence for goal alignment, stable contracts, and real risks. TDD is selected inside Engineering when a stable expected behavior makes test-first useful—it is not the lifecycle for every change.

GitHub stores recoverable goals, evidence, decisions, and delivery links. Issues can move backward to exploration, stop without code, or compress phases when the work is deterministic.

Projects can commit .devkit/project.json to provide reusable skills with project-specific facts without hardcoding them into Devkit. The profile can declare native commands, canonical workspaces, development-process rules, testing/staging/production targets, health checks, approved statistics access, and reporting cadence.

Copy skills/using-devkit/assets/project.example.json as a starting point. Profiles contain defaults, not authorization, and must never contain passwords, tokens, private keys, or other secrets.

A comprehensive Python code cleanup skill that enforces:

• Pyright strict mode with proper venv configuration
• Modern Python 3.12+ — type statements, built-in generics, X | None
• Automated formatting — ruff format, ruff check --select I
• Full type annotations — parameters, returns, instance attributes
• Clean imports — no empty __init__.py, no TYPE_CHECKING, no re-exports
• Pylint fix patterns — unused arguments, broad exceptions, import ordering
• TypedDict / Protocol / Literal over loose dict / Any

A prose cleanup skill for Chinese and English writing that:

• Routes by context — adapts everyday writing, technical documentation, and PR descriptions
• Preserves facts and protected spans — keeps names, numbers, commands, code, and links intact
• Calibrates to author samples — matches the writer's existing voice when examples are provided
• Protects against false positives — avoids changing clear, intentional, or already-natural phrasing

Web design guidance that reads the product, audience, brand assets, existing stack, and constraints before choosing a visual direction. It covers redesign audits, responsive behavior, accessibility, real UI states, asset rights, and rendered verification without forcing React/Next or a stock “AI” aesthetic.

An implementation workflow for screenshots, Figma frames, and approved visual references. It inventories layout and visual tokens before coding, keeps the existing project stack, and verifies fidelity at the source viewport plus responsive variants.

Creates image-only website references. Image count follows the requested sections, frames share one visual system, and each output stays readable enough for downstream implementation.

Creates image-only native mobile screens and flows with platform-aware safe areas, navigation, touch geometry, readable type, and consistent product states. Frame-free implementation references are supported.

A clean-room PowerPoint workflow for reading, creating, and safely editing .pptx/.potx files. It preserves the source, preflights available libraries and renderers, checks package/content integrity, and distinguishes completed visual QA from unverified rendering. Repository-backed reporting adds real git/tracker provenance, previous-plan comparison, stable metric definitions, diffable report.md + data.json sources, recurring archives, and weekly, monthly, architecture, MVP, release, or project-status narrative arcs.

A target-aware operations workflow for planning, deployment, local runtime startup, diagnosis, rollback, and hardening. It reads project-native contracts and optional project-profile environments, starts with bounded read-only evidence, distinguishes liveness from readiness, changes the smallest coherent unit, and repeats the original probe before declaring recovery.

A narrow production reporting workflow separated from deployment and development. It requires explicit production-statistics intent, a dedicated read-only account, bounded aggregate queries, database-native read-only transactions and timeouts, sanitized tables/charts, and transparent metric definitions. DDL, DML, repair, raw PII exports, and administrator shortcuts are outside the skill.

A complete user-space proxy installer and manager:

• Mihomo (Clash.Meta) binary — auto-detects OS and architecture
• Clash subscription integration with auto-update (every 30 min)
• MetaCubeXD Web UI panel for managing proxy rules
• Proxy wrappers — with-proxy, proxy-agent, proxy-copilot, proxy-claude
• Service management — systemd (Linux) / launchd (macOS) user services
• Uninstall — full cleanup of all files, services, and configuration
• No root required — everything runs in user space

Selects the smallest useful Issue profile for the current decision:

• Goal/Product — outcome, success signals, non-goals, and evaluation
• Exploration — hypotheses, evidence plan, budget, and stop condition
• Prototype — minimal artifact boundary and promotion/disposal rule
• Engineering/Bug — selected change, risks, contracts, and verification map
• Learning — evidence-backed proposal for reusable system improvement

It preserves safe gh issue create --body-file execution without forcing every kind of work into one template or adding a duplicate preview after direct authorization.

A phase-aware controller for existing Issues:

• recovers Goal and Evaluation Contracts from SIES or legacy Issue bodies
• resumes from the first unsupported decision instead of replaying a workflow
• allows Evaluation to return to Exploration/Prototype or stop without code
• chooses branches, worktrees, review depth, and tests from actual risk
• preserves trusted Issue targeting, body-file safety, traceability, and one delivery regression

Builds product Goal Contracts from known evidence and asks only questions whose answers can change outcome, scope, evaluation, or irreversible impact. Success signals may use user scenarios, metrics, visual results, contracts, or tests; “all tests pass” is not a product outcome.

skills/
  using-devkit/
    SKILL.md                    # Unified discovery and repository entry
    assets/
      project.example.json      # Copyable project parameter template
    references/
      level-decision.md         # L0-L3 depth scaling
      language-stack.md         # Language/tool detection
      project-profile.md        # .devkit/project.json discovery and schema
      orchestration-cheatsheet.md
      system-learning.md
  sies-engineering/
    SKILL.md                    # Goal-first engineering lifecycle
    references/
      evidence-strategy.md      # Evaluation, TDD selection, regression
      github-state.md           # GitHub Goal/Evidence state
  friendly-python/
    SKILL.md                    # Main skill document
    references/
      fix-patterns.md           # Before/after fix examples
      tool-codes.md             # Pyright/Pylint code reference
      advanced-patterns.md      # Complex type patterns
  humanizing-writing/
    SKILL.md                    # Chinese/English writing workflow
    references/
      chinese-patterns.md
      english-patterns.md
  taste-skill/
    SKILL.md                    # Context-aware web design
  image-to-code-skill/
    SKILL.md                    # Visual reference to frontend
  imagegen-frontend-web/
    SKILL.md                    # Image-only web references
  imagegen-frontend-mobile/
    SKILL.md                    # Image-only mobile flows
  pptx/
    SKILL.md                    # PowerPoint artifact workflow
    references/
      repository-reporting.md   # Traceable weekly/monthly/project reporting
  deployment-operations/
    SKILL.md                    # Deploy, diagnose, verify, and roll back
    references/
      deployment-playbook.md
      incident-response.md
  production-statistics/
    SKILL.md                    # Safe production aggregate statistics
    references/
      read-only-database.md
  mihomo-proxy-setup/
    SKILL.md                    # Installation/uninstall flow
    references/
      config-templates.md       # Config files and script templates
      systemd-units.md          # Linux systemd service units
      launchd-plist.md          # macOS launchd plist files
      troubleshooting.md        # Common issues and fixes
  github-create-issue/
    SKILL.md                    # Adaptive Issue creation
    references/
      issue-profiles.md
  github-issue-workflow/
    SKILL.md                    # Phase-aware Issue controller
    references/
      issue-sources.md
  github-product-manager/
    SKILL.md                    # Product Goal discovery
    references/
      question-framework.md
      issue-template.md

Re-run the install script — it auto-detects existing installations and updates via git pull:

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/zhpeng24/devkit/main/scripts/install.sh | bash -s -- claude

# Windows
& ([scriptblock]::Create((irm https://raw.githubusercontent.com/zhpeng24/devkit/main/scripts/install.ps1))) -Platform claude

# Claude Code
claude plugins update devkit

# Cursor
cd ~/.cursor/plugins/local/devkit && git pull

# Copilot CLI
cd ~/.copilot/plugins/devkit && git pull

# Codex
cd ~/.codex/devkit && git pull

# OpenCode — restart OpenCode (auto-updates from git)

# Gemini CLI
gemini extensions update devkit

1. Fork the repository
2. Create your skill in skills/<skill-name>/SKILL.md
3. Submit a PR

MIT License — see LICENSE. Adapted components retain their original licenses and attribution in THIRD_PARTY_NOTICES.md.