mateusands

onboard-agent

Onboard a new CLI coding agent into the crew as an audited MCP server, after a containment battery it must pass first.

plan

Plan a fix or feature before any code — acceptance criteria, blocking questions, falsifiable assumptions, blast radius, test plan, rollback point. Then stop.

Server conventions — layers and where each rule lives, input validation, errors through one path only, authorization and scoping, pagination, transactions and idempotency. Researches the docs for the exact version when a library's API is uncertain, requires TDD and validates green. Use when touching a route, handler, service or data access.

compliance

Compliance gate — personal data, dependency licenses, new external services and confidentiality. Run it BEFORE changing anything that touches sensitive data, adding a dependency or integrating a third party. Returns a verdict with the stopping point.

design-review

Visual craft review of UI changes — hierarchy, typography, color, depth, states and motion, with severity and a false-positive filter. Complements codereview (which does not look at the visuals). Use before opening a UI PR, or when asked to "review the screen" / "it looks generic".

frontend

Interface conventions — which design system to use, where types and data come from, mandatory states, theming, permissions as UX rather than as a barrier. Researches the docs for the exact version when a library's API is uncertain, requires TDD and validates in the browser. Use when touching anything visual or stateful.

local-environment

Bring up, seed and drive the local development environment — installation, platform-specific dependencies, stand-ins for external services, test data and the wiring traps that cost the most time. Use when running, manually testing or debugging the environment.

plan-review

Adversarial review of a plan written by ANOTHER agent, before any code exists — checks that the problem is the real one, that the assumptions are falsifiable and actually verified, that the blast radius was grepped rather than guessed, that the twin was hunted, that the test plan can fail, and that the boundary is drawn. Returns a verdict. Use when Claude wrote a plan and Codex reviews it, or the reverse. Never review your own plan.

schema

Database changes — the real shape of the schema, the "did it get applied in the right place?" gotcha, compatibility with the old code during deploy, column conventions and what is never deleted. Use when creating or altering a table, column, index or trigger.

comments

What to write in a code comment and what to delete — the test for whether a comment earns its place, the four kinds that do, the two red flags that mean it does not, and TODO conventions. Use while writing or reviewing any comment, and when a diff adds comments that restate the code.

coder

Execution of an approved plan — the discipline of the ACT of writing code, between the plan skill (before) and codereview (after). Confirms the right layer, runs the twin protocol, forces Red→Green→Refactor with proof, applies the conventions review will demand, self-reviews its own diff and STOPS when an assumption falls. Use when starting to implement any already-planned and approved fix/feature.

codereview

Senior code review of the repository's latest changes (working tree or latest commits) — intent before diff, triage by trigger, scalability/maintainability/security pillars, precision before coverage, severity and verdict. Only reports problems with file/line and the suggested refactor; does not apply fixes without an explicit order.

complete-security-review

Full-repository security audit — sweeps entire classes of problem (secrets, authorization, injection, upload, crypto, dependencies, infrastructure) across the whole codebase rather than a diff, states coverage as numbers, classifies by exploitability, and ends by running your host's own security reviewer as an independent pass. Use for a periodic audit, before a first release, when inheriting a codebase, or when asked to "check the security of the project".

end-session

Closes the working session — writes the permanent hardening report to .crew/hardenings/, updates the source of truth and the .crew memory files if anything they claim has changed, runs the final validation and honestly reports what was left out. Use at the end of every session.

local-testing

Tiered validation (L1 suite+build · L2 the production artifact opens · L3 real flow with real data · L4 hostile environment · L5 the artifact is actually deployed where you are judging it) of any change, with the recipe for computing the expected value before looking at the screen and the traps that have already produced "false green". Use before saying any change is done.

start-session

Initializes the working session — reads the project's source of truth, the .crew memory (techstack, operations, hardenings), the real git state and the loose ends from the last session, in READ-ONLY mode, creates the memory files if missing, expires old plans, and confirms scope alignment and active gates before any code. Use at the start of every session.

diagnosing-bugs

Find the cause of something that is broken, slow, flaky or intermittent — build a loop that goes red before forming any theory, rank falsifiable hypotheses, change one variable at a time, and end with a regression test or a written reason there is none. Use whenever behaviour is wrong and the cause is not yet known.

writing-for-agents

How to write a document an agent will run — a skill, an AGENTS.md, a workflow, a prompt. Covers what earns permanent context, where material sits between inline and disclosed, how a step's completion criterion drives thoroughness, and why a prohibition activates what it forbids. Use when creating or editing any of them.