README.md

software-development

Atomic, single-purpose skills for building software with an agent. Most are grouped by phase (design → … → verification); the stack/ group is orthogonal — knowledge tied to the tech you build on, not a phase. Each skill does one job and is unaware of the others — sequencing lives in workflows/, not here. Sources for adapted skills are in ../CREDITS.md.

design — shape intent into a design

Skill	What it does
brainstorming	Turn a brief or vague idea into an approved design doc via one-question-at-a-time dialogue.
domain-modeling	Actively build and sharpen the project's domain model — challenge terms against the glossary, sharpen fuzzy language, and write CONTEXT.md + ADRs inline as decisions crystallise. Composed by the grill-with-docs workflow.

planning — turn a design/spec into an actionable plan

Skill	What it does
writing-plans	Expand a spec into a bite-sized, TDD-shaped implementation plan with exact files, code, and commands.

review — stress-test a plan, or clean up / evaluate produced code

Skill	What it does
scrutinize	Outsider end-to-end review of a produced PR/diff/design doc: question intent → trace the real code path → verify the claim → severity-ordered findings + one verdict. Read-only (hands off edits to simplify/slop-cleanup).
receiving-code-review	Evaluate review feedback with rigor — verify each claim, push back when wrong, implement what holds up.
slop-cleanup	Detect and remove characteristic AI-generated slop from a diff, behavior-preserving.

engineering — write, debug, and test code well

Skill	What it does
karpathy-guidelines	Behavioral guidelines to reduce common LLM coding mistakes.
diagnose	A feedback-loop-first loop for hard bugs and perf regressions: reproduce → minimise → hypothesise → instrument → fix → regression-test.
post-mortem	Write the canonical record of a fixed bug (or resolved incident) — root cause, mechanism, fix, validation, how it slipped through. Refuses until the fix is validated; blameless. Pairs with diagnose.
de-flaking-tests	Make flaky tests deterministic (condition-based waiting) and kill tests that pass for the wrong reason (mock theater, incomplete mocks).
git-commit	Turn a working tree into clean, atomic, bisect-safe commits — Conventional-Commit messages, no co-author trailer, push left to the user.
git-worktree	Create/enter an isolated feature worktree and bootstrap-or-surface its environment (setup), then remove/prune it (teardown). Use when starting or wrapping up isolated agentic work.
delegate-coding	When the plan is clear enough for a cheaper agent to execute, delegate the coding to a headless executor CLI — cursor-agent, codex, or a cheaper claude — while your expensive "brain" model only plans, verifies, and owns the merge. Executor self-loops on env-independent checks in its worktree; you own env-dependent checks post-merge; bounded retries, then you finish. Per-tool invocation in references/.

verification — prove work is actually done

Skill	What it does
verification-before-completion	Gate before any "done / fixed / passing" claim: run the real command, read the output + exit code, then claim.

stack — specific to the tech you work in

Orthogonal to the phase groups above: knowledge tied to the technologies you build on (LLM/AI, frameworks, tooling) rather than a general SWE phase. Prefer durable playbooks (principles + links to official docs) over version-pinned API references, which rot.

Skill	What it does
llm-cost-optimization	Cut token/$ cost of LLM pipelines: measure → cache → tier → trim → batch → cap → verify.