AI workflows, for the whole engineering loop.
Prompts, skills, commands, configs, and agent patterns for planning, coding, reviewing, debugging, documenting, and shipping with verification close to the code.
The CLAUDE.md that keeps agents grounded
A short project memory file for the stack, invariants, and the lines an agent must not cross.
Remove trivial tests before they harden
An agent skill for replacing shallow AI-written coverage with tests that can catch real regressions.
Verification gate before done
Run typecheck, lint, tests, API flow, and build; keep fixing until the repo agrees.
Write tests that can actually fail
Behavior checks, negative cases, regression coverage, and cleanup that keeps the next run honest.
Security review of the actual diff
A read-only pass over dependency bumps, auth changes, and CI edits, with a fixed checklist and findings format.
UI verification with real browser evidence
Run user and guest flows, include hard refreshes, and keep screenshots as proof of what changed.
End-of-task reports that are useful later
A short closeout with changed files, commands and results, decisions made, and honest follow-up.
Root-cause investigation before edits
Find the exact source with file:line, name the systemic cause, and avoid masking symptoms with a patch.
Plan before touching the code
For large or ambiguous work: inspect first, write the steps, ask the blocking questions, then execute.
PHP to Python with golden-file parity
Move behavior across languages with fixture proof, uv tooling, and boring first-party libraries where they fit.
Final cleanup pass before calling it done
Re-scan the whole project for leftovers, stale docs, missed simplifications, and rule violations.
Parallel agents with clean write scopes
Plan first, split the work into disjoint surfaces, fan out, then reconcile and verify the whole result.
Commit, push, and open the PR cleanly
Branch when needed, make one conventional commit, keep the PR scoped, and report the exact SHA.
Observability that makes failures inspectable
Traceable logs, level control, test ids where they help, and no silent failure paths.
Read-only audit with file:line findings
A strict no-write pass that reports one concrete finding per line, sorted by impact.
Prove parity during a refactor
Capture reference behavior, separate public from internal fields, and make the new path match what matters.
Fix the whole pattern, not one instance
Find every occurrence, patch the full surface, and write the rule down so the drift does not come back.
Docs and diagrams before the work is done
Update the README, comments, Mermaid flows, and AGENTS.md decisions while the context is still fresh.
House rules agents can actually follow
The recurring repo rules — cleanup, stack, schemas, IDs, CI — written as constraints an agent can check.
DaisyUI and Tailwind without visual drift
Use the real components, match the reference theme, and verify every page in light and dark.
Commit and PR text from the real diff
A conventional commit subject and PR body that say what changed, why it matters, and how it was checked.
Compress a worklog without losing signal
Turn messy session notes into a dense developer journal while preserving decisions, commands, and evidence.
Add Discord notifications without breaking Slack
Find every Slack notification path, add the Discord twin, and keep the old channel working.
Domain and TLS expiry watcher
A daily check that ranks domains and certificates by expiry, then keeps one rolling issue current.
Dependency updates that only land green
A scheduled job bumps packages, runs the repo checks, and pushes only when the branch earns it.
AGENTS.md as the house-style file
One place for toolchain, code style, and repo invariants, with every other agent note pointing back to it.
Incident timelines while evidence is fresh
An agent follows logs and metrics during an incident so the timeline is accurate before context fades.
Teach agents to read EXPLAIN
Give it the plan, get back the slow node, estimate drift, and the index worth testing.
Staff-engineer review with a narrow charter
A terse review prompt for correctness, retry-safety, blast radius, and the smallest useful fix.
Postmortems from logs, not memory
Turn the incident channel and system evidence into a timeline that names systems, never people.
Spec, scaffold, then tests
Write the design first, let the agent scaffold the shape and failing tests, then fill in the logic.
Gate production migrations before review
Block locking DDL, unsafe deploy order, and unbatched backfills before a human has to catch them.
design.md for keeping generated UI on-system
The short token and component contract that stops agent-built UI from drifting into visual noise.