Standard feature
single packageA single-package feature with a manageable acceptance-criteria count.
feature-docACs → test list→designoptional→tddred·green·refactor→simplifytighten sweep→prod-readypre-merge→PR
AgentLead Skills · Claude Code · Codex · Antigravity · OpenCode
Pick by what’s in front of you.
AgentLead Skills encode engineering discipline so your assistant follows it by default. You don’t choose a skill — you describe the task and the right flow fires: investigate before building, design the interface first, test-first, review at the boundary, verify before merge.
doc — writes a contractshape — decides structurebuild — writes codegate — verifies before mergediagnose — analyses first
Where to start
Match your situation to an entry point. The flow takes it from there.
New feature in an existing systemfeature-docStandard feature →Large feature — many ACs or multi-packagefeature-doc + tdd-roundsLarge feature →New feature, direction still unclearinvestigateInvestigation →Bug — root cause obvious from the tracetddBug fix →Bug — intermittent, flaky, or cause uncleardebug → tddDebug flow →Refactor / improve existing architectureimprove-codebase-architectureRefactor →New multi-module system from scratchsystem-designGreenfield system →Reviewing someone else's PRpr-reviewReview →
Typo, dep bump, one-line configjust fix itno skill needed
The flows
Each is a chain of skills. Doc steps write a contract; gate steps refuse to advance until they pass.
Large feature
≥10 ACs · multi-packageOrchestrated across rounds — the parent dispatches Builder sub-agents and holds state.
feature-docmany ACs→tdd-roundsdesign·tdd·simplify per round→verify-real-depsreal upstreams→tag vN.0
Investigation before commitment
structural decisionFor a new dependency, framework choice, or a cross-cutting refactor — options before code.
investigate2–3 options→grill-planstress-test → ADR→feature-doccapture direction→→ feature flow
Bug fix — cause obvious
clear traceThe symptom and stack trace point straight at the bug.
tddfailing test → fix→prod-readyif infra / auth / DB→PR
Bug fix — cause unclear
flaky · intermittentEnvironment-specific, a regression, or you don't yet know what to assert.
debugreproduce → name cause→tddrepro = failing test→prod-readyif surface-changing→PR
Refactor existing code
brownfieldImprove architecture — turn shallow modules into deep ones. One candidate at a time.
zoom-outif unfamiliar→improve-codebase-architecturecandidate list→designdeepen interface→tddtests stay green→prod-ready
Greenfield system
once per systemA new multi-module project. After topology is set, each feature runs its own flow.
investigateoptional→system-designmodules · seams→designper module→feature-docfirst feature→→ feature flow
Review someone else's PR
utilityRuns whenever you read a diff — classifies findings by severity.
sync-checkdrift scan, optional→pr-reviewblocker · suggestion · nit
Lenses & gates that apply across
These don’t have a flow of their own — they hook into the ones above.
code-hygiene
lensBoring code, naming as the primary refactor, YAGNI, rule of 3, locality. Carried into simplify and pr-review.
security-review
gateFires on surface-changing work — new entry points, identity flows, authz, sensitive data, external deps. Runs alongside tdd.
zoom-out
utilityInterrupts any flow when you're lost in unfamiliar code. Maps the relevant modules and callers, then hands back.
bench
measurePerformance benchmarking discipline — measures latency and throughput when a flow carries a performance AC.
Install
One command. Works for Claude Code, Codex, Antigravity, and OpenCode.
# interactive — pick assistants & scope
$ npx @ai-agent-lead/skills
# everything, no prompts
$ npx @ai-agent-lead/skills --all --force
Scope it with --global (~/.claude/skills) or --local (./.claude/skills); target one assistant with --claude, --codex, --antigravity, or --opencode.