Claude Skills by Pixel-CLI

Operate real pi, agy, Claude, and Codex sessions in Herdr to diagnose tool behavior from transcripts, ask a CLI directly when its own behavior is unclear, then implement, retest, and redeploy a verified fix.

Replay Pixel's new-user setup in Docker without an LLM — install through a release archive, install.sh, the Homebrew tap, or sources from main, a PR head or a pinned commit; then personal Claude/Codex/pi settings kept, idempotent reinstall, first pixel audit, doctor and --fix, project install, disabled classify and uninstall; with --agents, Claude Code, Codex and pi sessions against a scripted fake model show whether each receives the Pixel prompt, runs pixel and routes grep. Use for Linux se...

Find deepening opportunities in pixel's crates (shallow modules to merge behind a smaller interface, leaking seams, logic only testable past its interface), scouted with pixel's own churn, size, caller and area data, presented as a local HTML report of before/after cards with their real cost (diff size, mutants, ARCHITECTURE.md sections, PR stack), then grilled through the candidate the user picks until it becomes a task on project 3. Adapted from Matt Pocock's improve-codebase-architecture s...

Mine the last 24 h (or a given window) of pixel usage — every repo's `.pixel/actions.jsonl` plus the agent transcripts indexed by `pixel recall` (claude, codex, pi…) — for frictions pixel actually caused, verify each one against the current binary and code, and propose ranked, evidence-backed improvements (bug fix, error message, flag, agent prompt, rule, doc); for the pixel repo itself, also measure each pull request from first edit to validated CI and where agents sat idle (`lead_time.py`)....

Deterministic code retrieval: indexed search, concept resolve, impact analysis, caller/callee tracing, task scoping, plan generation, and git history archaeology via the `pixel` CLI. Use when the repo has a `.pixel` directory, when the user mentions pixel, or before editing a symbol when blast radius matters. Requires the `pixel` binary on PATH.

Cut a pixel release end to end — pick the version, prepare the release commit (changelog cut, lockstep version bump, Cargo.lock, check-release) with prepare.sh in a pull request into main, tag vX.Y.Z on its merge to trigger .github/workflows/release.yml, and verify the published release, the install.sh asset and the Homebrew tap with evidence. Also covers resuming an interrupted release, patch releases while main is not releasable, and recovering from a failed Release run. Use when asked to r...

Microsoft's Pragmatic Rust Guidelines (M-* rules) applied to new or changed Rust code in this workspace. Use whenever writing, refactoring or reviewing code under crates/ — API shape, error types, panics vs errors, docs, performance, test design. Adds the borrowing, type-state, TODO and test-naming rules from Apollo GraphQL's rust-best-practices that Microsoft does not cover. Complements .agents/rules (mutation gate, lint idioms), does not replace them.

Redesign one or two sections of the pixel-cli.dev home (website/layouts/index.html) with the loop the hero went through for task #401 — critique as a visitor, an artifact of 2–3 mocked variants with votable points, the user's vote, synthesis iterations, every claim checked against the code and the benchmarks, then the Hugo implementation with desktop and mobile captures. Use when the user says "/section-redesign", "on itère sur la section …", "refais la section #problem", "même méthode que le...

Publish Pixel PRs promptly with optional local diagnosis and required CI validation.

Write or revise a pixel pull request body: fill .github/pull_request_template.md (Closes #N, Summary with a call-tree, file-tree or pseudocode sketch built from pixel's own graph, before/after Evidence with commands, Merge Danger as door and blast radius, Docs touched). Use when opening a PR, rewriting its description, or answering a reviewer who asks what a PR does or risks. Adapted from Matt Pocock's MIT pr skill and HumanLayer's MIT show-me skill (UPSTREAM pins both).

Make fast bounded classification decisions through `pixel classify` — labels plus per-label criteria in, a calibrated probability distribution and confidence out. Use when a task or a harness needs a typed judgment — intent routing, risk scoring, yes/no gates, triage, severity grading — without writing Jev integration code. Also use when building or improving an agent harness that should route, gate or grade on cheap model verdicts.

Trace a named code entry to a named destination when the request asks how execution reaches that point; exclude generic explanations, text searches, and edits.