robanderson

Use when a plan, design, spec, RFC, ADR, architecture, or technical decision needs to be rigorously stress-tested BEFORE any code is written. Basic triggers: user says "grill me", "grill this plan", "interview me about this", "poke holes in this", "pressure-test"/"stress-test"/"red-team"/"interrogate" my plan/design, "pre-mortem this decision", "what am I missing", "challenge my assumptions", "vet"/"sanity-check" this approach, "find the gaps/holes before I build", or hands over a proposal wanting relentless one-question-at-a-time Socratic questioning to reach a shared understanding. Hard opt-in triggers: a trailing "grill me hard" marker — "grill me hard", "grill hard", "hard grill", "grill me hard on the outstanding questions". Also use when a plan reads vague or risky. Planning-only: it never writes implementation code. NOT for reviewing already-written code (use a code-review skill), NOT for executing an agreed plan.

Set up an automated AI pull-request reviewer on Gitea or Forgejo Actions that runs Claude Code (headless `claude -p`) against a GLM model via the Z.AI Anthropic-compatible endpoint, posts the review as a PR comment, and sets a pass/fail status check that can gate merges. Use this whenever someone wants automated PR review / AI code review in CI on a self-hosted git server (Gitea, Forgejo, or a Gitea-family Actions runner), wants to wire Claude Code or GLM/Z.AI into their Actions pipeline, asks for a "review bot" or "Claude reviews my PRs", or is debugging an existing Claude-Code-on-Actions reviewer (529 overloads, runner not picking up jobs, comment/status not posting, the job dying silently). Prefer this skill over hand-rolling a workflow even if the user only says "review my pull requests automatically".

Run a Best of N tournament in one of two modes. First ask the user which model quality to use for the attempts (Opus, Sonnet, Haiku, or Mixed). SINGLE PASS: produce N independent solutions in parallel, then a blind Opus reviewer scores them, lists pros and cons, ranks them, and names a winner. TWO PASS: the same first round, but the Opus reviewer also distils what worked and what failed into guidance; the losing attempts are discarded, a second round of N fresh attempts is run with that guidance (positives to emulate, pitfalls to avoid), the saved round one winner is added back, and a final Opus ranker picks the overall winner. Trigger on a suffix like 'Best of N: 4' (single pass) or a two-pass marker such as 'Two Pass Best of N: 4', '2-pass best of n', or 'Best of N: 4 two pass' (two pass), case-insensitive; also trigger whenever the user describes a generate-and-rank tournament, with or without an intermediate learning round, even without the exact phrase.

Use Mockoon CLI to create API proxies with full request/response logging, mock API servers for testing, and record real API interactions for debugging. Use when testing APIs, creating compatibility layers, debugging microservices, analyzing API traffic patterns, or when the user mentions Mockoon, API mocking, mock servers, or API proxies.

Standardize Python development using UV - a fast, unified package and project manager. Use when creating Python projects, managing dependencies, setting up virtual environments, installing Python versions, or optimizing Python workflows. Replaces pip, virtualenv, pyenv, poetry, and pipx with a single 10-100x faster tool.

Run a Joust Engine tournament in one of two modes. The sigil is @@JE[:N][:M[:Z]] — N (optional) = attempts per round, M = passes (1 single, 2 two), Z = grand loops (Z>=2 = an UNATTENDED chain that, per loop, runs a full tournament, implements the winning proposal into your real repo on a new JE-<loop>-<random7> branch, runs fail-closed verify, and opens one PR — never auto-merged; Z=1 or omitted = today's isolated tournament; Z capped at Z_MAX=5); N may be inferred from a prose model spec like '2 opus, 2 glm 5.2, 1 codex high' (sum of counts = N, the items become the per-attempt Mixed assignment) or the Top Mixed preset ('top mixed' + N spread over opus/glm-5.2/codex-high), and bare @@JE falls back to the interactive model gate. A marker-adjacent task-size word (short/medium/long, e.g. @@JE:5 long) optionally overrides the per-attempt turn/timeout limits, which otherwise scale to the orchestrator's estimate of task size. First ask the user which model quality to use for the attempts (Anthropic Opus, Sonnet, H

Benchmark generation throughput (cold vs hot tok/s) for every model the joust-engine system can call (Anthropic / GLM / local MLX / codex / MiniMax / claudex via a local CLIProxyAPI proxy). Two workload profiles — light (tiny paragraph) and heavy (>5k-token input context + long >5k-token output, representative of coding/agentic work). Thin wrapper over bin/je-bench.mjs. Use when the user asks to benchmark model speed, measure tokens/second, compare cold vs hot throughput across providers, or run /je-bench.

Fable Engine (@@FE[:N]) — the fast composer variant of the Joust Engine tournament. One WIDE parallel round of N draft plans (mixed providers, blind-staged by the engine's composeOnly mode), then the ORCHESTRATING model itself reviews every draft at once, composes the best composite plan (crediting what it took from each candidate, discarding the rest, adding its own ideas), implements it with unit tests, and reports. No judging councils, no second round, no steelman loop — a ~30-minute turnaround instead of hours. Designed for a frontier-class orchestrator (Fable/Opus tier); it trades the blind council's independent ranking and second security gate for speed, so keep @@JE runs as the periodic calibration baseline. Trigger on the sigil @@FE[:N] (e.g. @@FE, @@FE:8), case-insensitive, optional marker-adjacent task-size word (short/medium/long).

Run a Farnsworth Loop tournament in one of two modes. The sigil is @@FL[:N][:M[:Z]] — N (optional) = attempts per round, M = passes (1 single, 2 two), Z = grand loops (Z>=2 = an UNATTENDED chain that, per loop, runs a full tournament, implements the winning proposal into your real repo on a new FL-<loop>-<random7> branch, runs fail-closed verify, and opens one PR — never auto-merged; Z=1 or omitted = today's isolated tournament; Z capped at Z_MAX=5); N may be inferred from a prose model spec like '2 opus, 2 glm 5.2, 1 codex high' (sum of counts = N, the items become the per-attempt Mixed assignment) or the Top Mixed preset ('top mixed' + N spread over opus/glm-5.2/codex-high), and bare @@FL falls back to the interactive model gate. A marker-adjacent task-size word (short/medium/long, e.g. @@FL:5 long) optionally overrides the per-attempt turn/timeout limits, which otherwise scale to the orchestrator's estimate of task size. First ask the user which model quality to use for the attempts (Anthropic Opus, Sonnet

Benchmark generation throughput (cold vs hot tok/s) for every model the farnsworth-loop system can call (Anthropic / GLM / local MLX / codex / MiniMax). Two workload profiles — light (tiny paragraph) and heavy (>5k-token input context + long >5k-token output, representative of coding/agentic work). Thin wrapper over bin/fl-bench.mjs. Use when the user asks to benchmark model speed, measure tokens/second, compare cold vs hot throughput across providers, or run /fl-bench.