Skip to content

gradigit/claude-skills

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

claude-skills

Opinionated workflow tools for Claude Code that I built and refined through daily use. They handle session management, project documentation, and skill development. They work best as a set, but individual skills can be used standalone where noted.

Installation

One-liner

bash <(curl -fsSL https://raw.githubusercontent.com/gradigit/claude-skills/main/install.sh)

Install everything non-interactively:

bash <(curl -fsSL https://raw.githubusercontent.com/gradigit/claude-skills/main/install.sh) --all

AI-guided install

Copy this prompt into Claude Code, Codex, or any AI coding agent:

Install skills from https://github.com/gradigit/claude-skills — read INSTALL.md and follow the install wizard.

The agent will walk you through language selection, skill picking, and installation interactively. Full wizard steps: INSTALL.md

Manual install

Option 1: Clone into a project (project-level)

# Clone into your project — skills are auto-discovered
git clone https://github.com/gradigit/claude-skills.git /tmp/claude-skills
mkdir -p your-project/.claude/skills your-project/.claude/commands
cp -r /tmp/claude-skills/.claude/skills/* your-project/.claude/skills/
cp -r /tmp/claude-skills/.claude/commands/* your-project/.claude/commands/

Option 2: Install globally (all projects)

git clone https://github.com/gradigit/claude-skills.git /tmp/claude-skills
mkdir -p ~/.claude/skills ~/.claude/commands
cp -r /tmp/claude-skills/.claude/skills/* ~/.claude/skills/
cp -r /tmp/claude-skills/.claude/commands/* ~/.claude/commands/

Option 3: Use as additional directory

git clone https://github.com/gradigit/claude-skills.git ~/claude-skills
# Then in Claude Code:
claude --add-dir ~/claude-skills

Skills

Meta / Skill Tooling

Skill Version Description
creating-skills 4.0.0 Creates new skills through adaptive interviewing or autonomous generation
updating-skills 2.0.0 Systematic workflow for modifying skills with version bumps and audits
auditing-skills 2.0.0 Validates skills against the Agent Skills spec and best practices
testing-skills 1.0.0 Three-tier evaluation of skills (triggering, functional, performance)

These four form a lifecycle: create a skill, update it as you iterate, audit it for quality, test it with evaluations. updating-skills references creating-skills for spec compliance and auditing-skills for the validation checklist. testing-skills runs EVALUATIONS.md scenarios to verify skill behavior.

Session Management

Skill Version Description
handoff 2.5.0 Creates HANDOFF.md so new sessions can pick up where the last one left off (defaults to local git-ignore for handoff artifacts)
handoff-fresh 1.9.0 Builds a fork-safe onboarding bundle for a brand-new agent in forked/new-folder repos (defaults to local git-ignore for bundle artifacts)
syncing-docs 2.6.0 Detects drift between code and project state files, fixes owned docs (CLAUDE.md + AGENTS.md)
managing-doc-manifest 1.0.0 Creates .doc-manifest.yaml — a registry of docs and their code references
wrap 1.3.0 End-of-session coordinator: sync docs, run instruction-doc quality pass (CLAUDE.md + AGENTS.md), then handoff (optional fresh bundle)

During work, syncing-docs keeps documentation in sync with code changes. At the end of a session, wrap chains sync-docs → claude-md-improver (Anthropic plugin) → handoff. For forked/new-folder continuation by a brand-new agent, run handoff-fresh to generate a full onboarding bundle.

Claude Code sessions are ephemeral, but project context shouldn't be. These skills maintain CLAUDE.md, AGENTS.md, TODO.md, HANDOFF.md, and .doc-manifest.yaml as persistent state files that survive /clear and crashes.

Explicit Commands and Hook Model

  • Canonical command entry points:
    • /sync-docs
    • /handoff
    • /handoff-fresh
    • /wrap
  • These flows are manual by default. No implicit side-channel behavior is required.
  • Hook guidance and archive-first cleanup policy for deprecated auto-handoff hooks: see HOOKS.md.

Research

Skill Version Description
study 1.2.0 Structured web research with source verification and hypothesis tracking

Standalone skill. No dependencies on other skills in this collection.

Forge (Autonomous Workflows)

Skill Layer Version Description
forge-claude-teams 0 (practices) 1.1.0 Claude Code best practices for Agent tool, TeamCreate, SendMessage
forge-codex-multiagent 0 (practices) 1.1.0 Codex CLI best practices for spawn_agent, fork_context, batch processing
forge-research 1 (capability) 1.1.0 Autonomous multi-agent research with hypothesis tracking and adversarial challenge
forge-builder 1 (capability) 1.2.0 Autonomous building/coding with self-review, self-improvement, and quality gates
forge-orchestrator 2 (orchestrator) 1.4.0 Sequences research and building through milestone-gated cycles with compound learning

Layered architecture with custom agents for execution isolation:

forge-orchestrator (Layer 2)
  ├── forge-research (Layer 1) ──→ forge-claude-teams OR forge-codex-multiagent (Layer 0)
  ├── forge-builder (Layer 1) ──→ forge-claude-teams OR forge-codex-multiagent (Layer 0)
  └── Custom Agents: forge-adversarial-reviewer, forge-build-worker,
                     forge-research-worker, forge-performance-auditor

Use /forge [goal] to launch the orchestrator, or invoke Layer 1 skills standalone with /forge-research [topic] or /forge-builder [task]. See AGENTS.md for the full dependency graph and custom agent details.

Why These Exist

Claude Code sessions are stateless — context is lost on /clear or crash. After enough sessions where I had to re-explain project state from scratch, I built these skills to fix three things:

  • handoff captures what was done, what failed, and what's next. A new session reads HANDOFF.md and picks up where the last one left off.
  • syncing-docs detects when code changes make docs stale and fixes them. No more outdated README sections or wrong file paths in CLAUDE.md.
  • The meta-tooling skills (creating-skills, updating-skills, auditing-skills, testing-skills) enforce consistency so skills don't degrade as you iterate on them.

Dependencies

Between skills in this collection

wrap ──→ syncing-docs ──→ managing-doc-manifest
  │
  └──→ handoff
  └──→ handoff-fresh (optional via --with-fresh)

updating-skills ──→ creating-skills (spec rules)
                └──→ auditing-skills (validation checklist)
                └──→ testing-skills (run evaluations)

testing-skills ──→ creating-skills (EVALUATIONS.md format)

study (standalone)

forge-orchestrator ──→ forge-research (research phase)
                   └──→ forge-builder (build phase)
                   └──→ forge-claude-teams OR forge-codex-multiagent (platform practices)
                   └──→ forge-adversarial-reviewer, forge-build-worker,
                        forge-research-worker, forge-performance-auditor (custom agents)

forge-research ──→ forge-claude-teams OR forge-codex-multiagent (platform practices)
               └──→ study (optional, fallback to direct WebSearch)
               └──→ forge-adversarial-reviewer (optional, for adversarial challenge)

forge-builder ──→ forge-claude-teams OR forge-codex-multiagent (platform practices)

wrap chains three skills sequentially by default, with an optional fourth step (--with-fresh) for handoff-fresh. If a dependency isn't installed, that step can be skipped — but the full workflow works best with all session management skills present.

External dependencies

Dependency Required by Type How to get it
claude-md-improver wrap (step 2) Anthropic official plugin Install from anthropics/skills

Optional integrations

  • syncing-docs checks for an architect/ directory (created by forging-plans). If it doesn't exist, this is silently skipped.
  • handoff references CLAUDE.md, AGENTS.md, TODO.md, and architect/ files if present. None are required.
  • handoff and handoff-fresh default to --ignore-mode local, which writes HANDOFF.md and .handoff-fresh/ into .git/info/exclude (local-only ignore). Use --ignore-mode shared to write ignore entries to .gitignore, or --ignore-mode off to disable.
  • handoff-fresh emits a fresh-agent bundle at .handoff-fresh/current/ by default: claude.md, agents.md, todo.md, handoff.md, context.md, reports.md, artifacts.md, state.md, prior-plans.md, read-receipt.md, session-log-digest.md, session-log-chunk.md, handoff-everything.md. It also updates root HANDOFF.md as a bridge pointer to bundle handoff, supports an agent-internal --validate-read-gate preflight check before coding, enforces shared-context parity between bundle claude.md and agents.md, and keeps log continuity token-budgeted.

Customization

These skills are workflow opinions. They assume:

  • Projects should have CLAUDE.md and AGENTS.md for persistent cross-agent context
  • Sessions should end with a handoff for continuity
  • Documentation should be tracked with .doc-manifest.yaml
  • Skills should follow the Agent Skills spec

You're free to modify any of these to fit your workflow. The skills are MIT-licensed.

Common customizations

  • Don't want .doc-manifest.yaml? Remove managing-doc-manifest and the manifest step from syncing-docs
  • Don't use HANDOFF.md? Use syncing-docs standalone without wrap or handoff
  • Want a different HANDOFF.md format? Edit handoff/templates.md
  • Want different audit rules? Edit auditing-skills/checklist.md

Related

License

MIT — see LICENSE.

About

Battle-tested Claude Code agent skills for meta-tooling, session management, and structured research

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors