noahwc/quiver

Metric-driven experiment loops: define a single measurable metric, baseline it, then iterate one change at a time — measure, compare, keep (commit) or discard (revert). State lives in git and a log file, not in context. Use for optimization work: make X faster, reduce bundle size, improve the benchmark, cut token usage, tuning parameters, or any run-measure-compare loop.

Split uncommitted changes into semantic commits and push the current branch. Use when a working tree mixes unrelated changes and the user asks to commit, clean up, or split them into logical commits, or invoke manually: /commit-split. Never rewrites history or force-pushes; asks before pushing the default branch.

Mine session history for (1) repeated user corrections, proposed as When → Do habits that each need explicit human approval, and (2) repeated questions, lookups, file reads, and commands in the current repo, turned into agent-cli subcommands stored outside the repo. Use when the user asks to mine sessions, learn from feedback, review recurring corrections, or build/refresh agent-cli — or when you notice the user correcting the same behavior, or yourself repeating the same lookup, for the second time or more.

Use when: presenting large findings (research, investigations, audits, comparisons, multi-section reports). Write a standalone HTML summary optimised for scannability and open it automatically, instead of printing a wall of text in chat/terminal.

Run deterministic tooling (typecheck, linters, autofixers) FIRST — before any agentic file-by-file inspection — when asked to find or fix syntax errors, type errors, code quality issues, any usage, or to check/clean up/audit the codebase. Two stages: (1) discover and use the repo's existing typecheck/lint/format setup; (2) if none exists, propose adding one before doing any manual inspection. Manual reading is the LAST resort, reserved for what tooling cannot catch or fix.

Stewards PRs authored by the current user (or explicitly taken over) through CI and review: triages unresolved Copilot and human threads first, treats a Copilot 'Needs a closer look' overview as a worklist, fixes failing checks or re-runs flaky jobs, and escalates architectural or breaking changes before editing. One bounded invocation — never waits for CI. Use when asked to fix CI or resolve review comments on your own GitHub PR, or when invoked as /pr-steward [#N].

Generate a short interactive quiz on the code the agent just wrote (or a diff), so you prove you understand the implementation before shipping it. Opens a self-contained HTML page with auto-graded questions about behavior, design rationale, failure modes, and change impact — not language trivia. Use after landing a non-trivial change, or invoke manually: /quiz-me. Complements slopchop-review (correctness) — this checks the human's understanding.

Hunk-by-hunk diff triage with three verdicts: FIX (must change, with the concrete fix), DISCUSS (question or tradeoff for the author), OK (silent). Ends with a FIX/DISCUSS summary and offers to apply the FIXes. For reviewing working-tree changes, staged changes, branches, or PRs. Invoke manually: /slopchop-review. Complements ponytail-review (over-engineering only) — this one covers correctness, security, and edge cases.