Reusable agent skills for development, ops, and productivity — battle-tested patterns with working examples.
Uses the universal Agent Skills specification. Compatible with Claude Code, Codex, OpenCode, Cursor, GitHub Copilot, Windsurf, and other Agent Skills hosts.
The minimal .claude-plugin/plugin.json is installer compatibility metadata: npx skills uses it to group the installed skills under Sean2077 Skills for every target agent.
Install only the workflows you expect an agent to discover. Every installed skill contributes routing metadata even though its full body, references, and scripts remain on demand.
# Recommended: install one skill for both Claude Code and Codex
npx skills add sean2077/skills --skill analyze -a claude-code -a codex
# Add a deterministic delivery workflow only where it is needed
npx skills add sean2077/skills --skill ralph -a claude-code -a codex
# Single-host examples
npx skills add sean2077/skills --skill code-review -a codex
npx skills add sean2077/skills --skill deep-interview -a claude-code
# Optional convenience: install the complete catalog
npx skills add sean2077/skills -a claude-code -a codex
# From a local checkout, keep the leading ./ so it is parsed as a path
npx skills add ./skills/agent-scaffold -a codexRepeat the selective command for each additional skill. If skills were installed before the compatibility manifest existed, rerun a catalog-root add command once in the same project/global scope so npx skills records the group in its lockfile. For a local checkout, use npx skills add . --skill agent-scaffold -a codex; installing ./skills/agent-scaffold directly bypasses the root manifest and therefore remains ungrouped.
| Skill | Description | Stack |
|---|---|---|
| agent-scaffold | Apply or refresh a dual-host (Claude Code + Codex) harness: .agents/ SSOT, mandatory real-symlink projections, merge-owned hooks, subagent projection, default/light governance profiles, and structured plan/doctor/verify output. |
Shell, Python, Governance |
| ai-slop-cleaner | Perform behavior-preserving, bounded cleanup of duplication, dead code, needless abstraction, boundary leaks, and weak coverage with explicit verification. | Engineering, Refactoring |
| analyze | Explain repository behavior or investigate failures through one read-only evidence workflow with ranked synthesis, competing hypotheses, falsification, and discriminating probes. | Engineering, Analysis |
| autopilot | Deliver authorized work end to end with a proportional native Agent loop by default and an opt-in persistent runtime for interruption-safe, cross-session, or audit-sensitive execution. | Python, Delivery |
| best-practice-research | Compare current primary sources and viable trade-offs to produce a repository-specific technical recommendation. | Research, Engineering |
| code-review | Perform an evidence-based defect review of a concrete change set, including regressions, security, maintainability, and verification gaps. | Engineering, Review |
| conventional-commit | Create one scoped local Conventional Commit or return one history-aware message-only subject while preserving unrelated index state. | Git |
| deep-interview | Turn vague ideas into an explicitly approved specification through an adaptive interview, with opt-in deterministic topology/scoring state for persistent or auditable sessions. | Python, Requirements |
| lark-cli | Route all 飞书/Feishu/Lark operations through one lean lark-cli entry point with on-demand domain references, explicit identity continuity, command discovery, and side-effect safety. |
Lark, CLI, Productivity |
| project-docs-organizer | Derive project-owned documentation structure from reader, task, domain, ownership, lifecycle, and retrieval evidence; use optional sibling-local numbering only when stable order materially improves reader navigation. | Documentation |
| prototype | Reduce one uncertainty through a disposable, bounded experiment with explicit oracle, safety, conclusion, and cleanup boundaries. | Engineering, Experimentation |
| ralph | Run a bounded verifier loop until it passes, stalls, plateaus, or exhausts its round budget, with deterministic state and compact receipts. | Python, Iteration |
| semver-release | Plan and publish a semver release with deterministic reachable-tag/bump analysis, a preferred changelog-backed tag workflow, project-owned version synchronization, and policy-derived publication verification. | Git, Python, Release |
| skill-eval | Run deterministic, comparable baseline/treatment evaluations for Agent Skills, including trigger leakage, verifier, changed-path, repository-isolation, and cost gates. | Python, Evaluation |
| tdd | Apply explicitly requested test-first implementation across stacks by deriving seams, oracles, test levels, and commands from the target project, with validated RED-GREEN-REFACTOR evidence and guidance for effects, legacy code, and hard cases. | Engineering, Testing |
| tooling-conventions | Derive project-owned command boundaries, placement, and evidence-gated safety contracts, with optional structural inventory reconciliation. | Shell, Governance |
| work-protocol | Externalize high-risk or cross-session tasks into durable artifacts with one loop-owner lease, CAS state, hash-chained evidence, and isolated writer/reviewer worktrees. | Python, Git, Coordination |
skills/
└── <name>/
├── SKILL.md # Single source of truth for every installer
├── references/ # Optional: category-named, on-demand detail (no catch-all document)
│ └── <category>.md
└── scripts/ / assets/ # Optional: deterministic helpers and output resources
scripts/
├── validate_skills.py # Catalog-wide frontmatter, name↔dir, README, reference, and placeholder checks
├── catalog_core.py # Shared error/warning collection and repository path constants
├── contracts/ # Optional targeted executable/high-risk contracts, discovered by filename
│ └── <skill>.py
├── workflow_runtime/ # Maintainer SSOT for common + workflow-specific deterministic runtime code
│ ├── common.py
│ ├── autopilot.py
│ ├── deep_interview.py
│ └── ralph.py
├── generate_workflow_runtimes.py # Generates/checks each independently installable single-file runtime
├── p0_runtime/ # Maintainer SSOT for skill-eval and work-protocol runtime packages
├── generate_p0_runtimes.py # Generates/checks independently installable P0 skill runtimes
├── test_validate_skills.py # Focused catalog-contract regression fixtures
├── tests/ # Behavioral and adversarial skill-specific regressions
├── check-agent-scaffold.sh # agent-scaffold static gate: syntax + install-depth invariant + dogfood drift
└── e2e-agent-scaffold.sh # agent-scaffold behavioral gate: install into a throwaway repo, assert it works
evals/examples/tdd/ # Offline positive/negative/confusable skill-eval example
evals/agent-skills/ # Live analyze/autopilot/deep-interview routing suites; CI validates manifests only
requirements-validation.txt # Pinned official skills-ref + StrictYAML validation dependency
.claude-plugin/
└── plugin.json # npx skills grouping metadata, kept in sync by the validator
.github/workflows/
└── validate.yml # Cross-platform catalog, runtime, install, shell, and behavioral gates
autopilot, deep-interview, and ralph each ship one generated, Python 3.8+ standard-library script. The autopilot and deep-interview runtimes are opt-in persistence/control planes; ordinary single-session delivery and interviewing stay model-native. ralph uses its runtime as the normal bounded verifier loop. Installing a single skill creates no sibling-skill, repository-runtime, or OMA CLI dependency. Maintainers edit scripts/workflow_runtime/, run python scripts/generate_workflow_runtimes.py, and let CI reject generated drift.
skill-eval and work-protocol likewise ship generated Python 3.8+ standard-library packages and remain independently installable. Maintainers edit scripts/p0_runtime/, run python scripts/generate_p0_runtimes.py, and validate the offline TDD A/B fixture plus the P0 behavioral/adversarial suites before committing.
Use model-native reasoning for reversible single-session work. Add deterministic scripts only around machine state or costly machine-checkable boundaries such as external side effects, Git/release integrity, concurrency/CAS, path and identity safety, generated drift, or comparable evaluation. Targeted per-skill contract modules are optional; the reviewed high-risk subset is registered so accidental deletion fails validation, while prompt-only semantics belong in SKILL.md and live evaluations rather than brittle phrase checks. See the harness constraint policy.
The former trace route is now the causal-investigation mode of analyze. Existing installations should replace trace with analyze and remove any stale trace projection to avoid duplicate routing.
The runtime scripts emit compact versioned receipts by default. Full state is opt-in through --full; history is bounded through history --tail; discovery is bounded through list --limit. Git repositories share discovery across worktrees while preserving explicit mutation ownership. Non-Git workspaces use a stable --root.
Reference filenames are descriptive lowercase kebab-case. Link every category directly under the resident SKILL.md On-demand references router, and state its conditional load boundary near the top. Do not add root-level reference.md files or catch-alls such as misc.md, all.md, or references/README.md.
npx skills reads directly from skills/, so this repository does not maintain separate .codex/skills or .claude/skills mirrors.
Run the catalog's pinned validation, official spec, discovery, deterministic runtime, shell, and behavioral gates from the development commands. Release-facing changes accumulate in the changelog. After the release snapshot is merged and validated on main, an annotated stable or numbered-prerelease v tag triggers the repository-owned release workflow. It reruns the complete validation workflow, extracts the exact matching changelog section from the tagged commit, and creates a GitHub Release only after both steps succeed.
MIT