Skills

These are agent skills: procedures an agent can follow for a particular kind of work.

There is no installer. A skill is a directory under skills/<name>/ with a SKILL.md file. Read that file. The name and description in the frontmatter say when the skill applies. The rest of the file is the procedure. A skill may also keep notes, scripts, or other files next to SKILL.md. Use those only when the procedure tells you to.

Optional personal profiles

Personal coordinator profiles and new-host setup are explicit opt-ins, separate from the reusable skills. They do not configure a host or log in to providers just by being present.

Setup

The skills in this repository do not need to be installed. Optional outside engineering skills can be downloaded with Git. Run this from the repository root (or invoke the script by its full path from another directory):

./scripts/setup-protocols.sh

That downloads two public collections into .external/, which git ignores. They are not part of the skills published here:

• mattpocock/skills lands in .external/mattpocock-skills/skills
• cursor/plugins (pstack/skills) lands in .external/cursor-plugins/pstack/skills

The script records both directories in .external/catalog.env. Load them into the shell that launches /delegate:

. ./.external/catalog.env

The file exports COORDINATOR_PROTOCOL_CATALOG, with the pstack catalog first. Paths containing spaces and shell metacharacters are quoted safely. Repository paths containing a colon or newline are rejected because the catalog uses : to separate paths.

Run the setup script again to fetch updates and fast-forward clean checkouts. Before changing either working tree, it checks both destinations and refuses unrelated existing directories, symlinks, a changed origin, detached or non-tracking branches, tracked edits, staged edits, untracked or ignored files, local commits that cannot fast-forward, and rewritten upstream history. It preserves those files and commits; move or preserve your work before retrying. It also refuses unrelated or edited catalog.env content rather than overwriting it. The exact catalog generated by the previous script version is upgraded automatically. Fetches can update Git metadata before a refusal, and a later Git/download error can leave a partially completed setup; rerun after resolving the error.

To use different sources, set MATT_SKILLS_URL and CURSOR_PLUGINS_URL to Git repository URLs or local repository paths before running setup. Relative local paths are resolved from the directory where you run setup; quote values containing spaces. Existing clones must still match the selected sources. The plugins source must contain pstack/skills, and the mattpocock source must contain skills.

If you skip setup, /delegate uses its normal fallback without external skills. To stop using downloaded catalogs in a shell that loaded them, run unset COORDINATOR_PROTOCOL_CATALOG before removing .external/.

The setup integration tests create real local Git repositories, run the actual script, source the catalog with POSIX sh, and invoke the existing Coordinator resolver. They block Git network protocols and do not change global Git configuration:

python3 -m unittest discover -s scripts -p 'test*.py'

Skills

• adversarial-review — Looks at one change or document, such as a pull request, spec, design, or plan, and reports concrete defects and gaps that the evidence supports.
• agentic-loop — Pursues delegated outcomes through planning, action, observation, verification, and adaptation, loading methodology only when needed.
• browser-qa — Verifies scoped user journeys and rendered behavior in a real browser, with reproducible evidence and explicit coverage limits.
• bug-retro — After a bug is fixed, turns that single fix into a lasting improvement for the whole class of bug.
• code-review — Inspects code changes for evidence-backed correctness, security, and compatibility risks.
• coordinator-setup — Discovers host capabilities and configures optional coordinator adapters in an external host profile without rewriting the base skill.
• debugging — Tests causal explanations through hypotheses, discriminating experiments, and model updates.
• deep-module-review — Judges one module or API for cohesion, what it hides, how heavy its interface is, what leaks out, and how much callers have to carry.
• delegate — User-invoked (/delegate). Splits substantial work across headless or native workers through a code-driven runtime that owns the flow and gates, and delivers Git changes or report/file artifacts. The agent answers typed judgments and performs requested native host calls.
• finding-triage — Decides which review findings must be fixed, what each fix would cost, and whether to route, narrow, accept, or reject it.
• herdr — Inspects and coordinates Herdr workspaces, tabs, panes, terminals, and agents from the command line without disrupting work already running.
• implementation — Makes controlled code changes while preserving contracts, surrounding conventions, and existing user work.
• invariant-reviewer — Looks for reachable failures and broken invariants in a design or implementation, and weighs how much complexity a proposed fix would add.
• investigation — Gathers traceable evidence to answer bounded engineering questions and reduce uncertainty.
• model-benchmark — Measures coding models and agent harnesses on real engineering tasks from the term2 repository, including solve rates and blind grading of candidate diffs.
• playwright-cli — Drives a browser to interact with web pages and to run or work with Playwright tests.
• profiling — Measures representative workloads, locates limiting mechanisms, and verifies optimizations with comparable evidence.
• proportionality-review — Judges whether a proposal or design is more complex than the problem needs.
• research — Sources, compares, and synthesizes evidence into traceable answers to bounded questions.
• risk-first-build — Starts non-trivial builds with concrete discovery, verifies assumptions in order of risk, and reports evidence honestly.
• simplicity-architect — Proposes software designs built on strong invariants and a small complexity budget, keeping edge cases from growing into extra defensive machinery.
• skill-creator — Creates and revises agent skills, and measures how well they perform, including tests and how accurately a description gets the skill selected.
• skill-router — Entry point for every request. Routes to exactly one exit: no skill (most requests), one skill, or a two-skill chain. scripts/check_router.py fails if a skill is missing from it, an eval expects an exit that doesn't exist, or a route exit has no eval; scripts/run_router_evals.py routes every eval through a model to catch rows shadowing each other.
• slop-audit — Surveys a whole repository for piled-up problems in design, correctness, tests, duplication, security, dependencies, and overbuilding, then ranks what to repair.
• supervise-panes — Runs and supervises hand-watched workers, such as agents in Herdr panes, through a phased procedure with a durable task record; also resumes older manual coordinator tasks. Uses the delegate skill's scripts.
• term2 — Starts, configures, and runs term2 for interactive coding sessions or one-shot command-line tasks.
• testing — Establishes proportionate behavioral evidence through meaningful assertions, isolated execution, and explicit test limits.
• workflow-evolution — Improves an agent workflow by revising its skills through experiments backed by evidence, leaving the harness unchanged.

Agentic Loop primitives

Use agentic-loop to own the delegated outcome. Load investigation, debugging, implementation, testing, code-review, research, browser-qa, and profiling only when the current phase needs their methodology. Compose them for the task instead of loading the entire set up front.

Each primitive includes optional source notes describing its methodological influences. Use the existing skill-creator when a real methodology gap remains; keep generated methods provisional and promote them only after repeated successful use, checking for overlap with existing skills.