Agent Skills for Real Engineering

Browse on skills.sh · Install from npm

Composable agent skills for engineering and productivity workflows. The fork tracks compatible improvements from mattpocock/skills, keeps its own local-first tracker and workflow adaptations, and works across Claude Code, Codex, OpenCode, and other skill-compatible agents.

Install

Pick one route. Both install the same promoted Engineering and Productivity skills.

Recommended — skills.sh

Use skills.sh with Claude Code, Codex, OpenCode, or another compatible agent:

npx skills@latest add ecology9191/skills

Pick the skills and coding agents you want. Include setup-agent-skills when you use the engineering workflow skills.

Fork package installer

The public npm package installs the complete promoted set or an explicit subset into ~/.agents/skills:

npx --package @ecology91/skills ecology91-skills

The interactive menu contains the promoted engineering and productivity buckets. Use --all, --bucket, or --skill for a non-interactive selection.

Run /setup-agent-skills once in each engineering repository, then restart the agent so it reloads its skill list.

Invocation model

User-invoked skills run only when a human explicitly names them. Model-invoked skills can also be selected automatically when their descriptions match the work. The Claude Code and Codex metadata enforce the same split; see the invocation contract.

Engineering

Daily code-work skills.

User-invoked

• ask-matt — Route a situation to the right skill or workflow in this repository.
• grill-with-docs — Stress-test a plan while maintaining the project domain model, CONTEXT.md, and ADRs.
• implement — Implement a spec or ticket set with agreed TDD seams and a closing code review.
• improve-codebase-architecture — Find high-leverage deep-module opportunities and work through the selected candidate.
• opensrc — Inspect package or repository source code as the hard source of truth.
• setup-agent-skills — Configure the local-first tracker, triage vocabulary, and domain-doc layout shared by the engineering flows.
• to-spec — Synthesize the current conversation into a spec and publish it to the configured tracker.
• to-tickets — Break a plan or spec into tracer-bullet tickets with explicit blocking edges.
• triage — Move issues and external merge requests through the repository's triage state machine.
• wayfinder — Map a multi-session effort as linked decision tickets and resolve the route before implementation.

Model-invoked

• code-review — Review a diff in parallel against repository standards and its originating spec.
• code-structure — Separate domain orchestration from shared operational mechanics when workflows duplicate behavior.
• codebase-design — Design deep modules with small interfaces, clean seams, and interface-level tests.
• diagnosing-bugs — Run a gated reproduce, minimise, hypothesise, instrument, fix, and regression-test loop.
• domain-modeling — Build and sharpen the project's shared domain language and architectural decisions.
• prototype — Build a focused, shareable prototype to answer a design question.
• research — Investigate a question from primary sources and capture a cited Markdown report.
• resolving-merge-conflicts — Resolve an active merge or rebase conflict hunk by hunk from both sides' intent.
• sandcastle-shadow-monitor — Supervise a scoped Sandcastle run while separating issue failures from reusable workflow churn.
• setup-coding-quality-checks — Install and document strict local quality checks for agent-assisted development.
• tdd — Build features and fixes as small red-green-refactor vertical slices.
• to-qa — Create a local QA To Do session from completed work under an explicit source issue.
• wizard — Generate an interactive shell guide for the steps only a human can perform.

Productivity

Daily non-code workflow skills.

User-invoked

• grill-me — Interview the user until every reachable branch of a plan or decision is resolved.
• handoff — Produce a portable continuation document for another agent, directory, or harness.
• teach — Teach a concept across sessions using the current directory as a stateful learning workspace.
• to-questionnaire — Turn a decision gap into a questionnaire for the person who can answer it.
• wait-what — Re-pitch the last message with missing context, plain language, and project vocabulary.

Model-invoked

• find-skills — Discover and evaluate installable skills for a concrete task or missing capability.
• grilling — Run the reusable frontier-based interview behind the grilling workflows.
• use-subagents-frequently — Delegate bounded independent work when breadth or context pressure warrants it.
• writing-for-agents — Write predictable skills, AGENTS.md, CLAUDE.md, and documents reached by agent pointers.

Repository layout

Only engineering/ and productivity/ are promoted and shipped. misc/ contains retained but unpromoted skills, in-progress/ is public beta work, and deprecated/ records retired skills. See each bucket's README for its complete contents.

For a local checkout, run scripts/link-skills.sh after adding, removing, or renaming a skill. It links every active skill into ~/.claude/skills and ~/.agents/skills, so future pulls update the installed copy in place.