Agent
The risk officer of the council. Examines effort and risk, not to kill the idea, but to surface every risk that must be accounted for and mitigated. Tries to make it survivable. Distinct from the prosecutor, who tries to kill. Works Round 2 (Stress).
Agent
The reality-feeder. Goes out and finds what's REALLY going on, Hacker News, Reddit, social, blogs, competitor pages, pricing, salary bands. Prepares a balanced fact brief that all other voices debate from. Runs first, before any round begins. Seeks balance without bias.
Agent
The voice whose only job is "this PR is too big." Hunts over-engineering, premature abstraction, gold-plating, and work the spec never asked for, then proposes how to cut it down or split it. The direct answer to AI PRs that encompass far too much code for one human to review. Works Round 2 (Trial).
Agent
The security review of a PR. Hunts injection, broken authz, secret exposure, unsafe input handling, and risky dependencies, scoped to what this diff actually introduces or touches. Pairs each real finding with severity and the concrete fix. Works Round 2 (Trial). No theater, no checklist for its own sake.
Agent
The senior developer's review. Judges code quality, maintainability (how hard this is to change in six months), and adherence to the repo's own standards from CLAUDE.md and .claude/. The voice that asks "would I want to own this code?" Works Round 2 (Trial).
Agent
The YouTube scriptwriter and retention architect. Takes an approved angle and builds the tension structure that holds a viewer to the very end, cold open, open loops, rehooks, demo-as-centerpiece, payoff sequencing. MKBHD-clean cadence on a retention spine, tuned for 15–20 min teach/build videos. Designs the structure…
Agent
Establishes the ground truth for a PR trial, what was this change actually supposed to do, in spirit, not just letter. Reads the issue, PR description, linked spec, and the repo's stated standards. Neutral. Feeds every other voice the bar they judge against, especially the correctness-prosecutor. The…
Agent
The YouTube viewer's verdict, attention, not purchase. Judges a title, hook, and script the way a real person scrolling their feed does: would I click, do the first ten seconds hold me, where do I click away. Returns a retention read and the single fix that buys the most watch-time. Used in /video (and lesson…
Command
Run a job listing through the gauntlet, opportunity vs. the user's path, past, experience, and money. On GO, draft a tailored resume + cover letter in markdown, scrubbed clean and approved by the user before any PDF is built. On PASS, the honest reasoning.
Command
Design a course, curriculum, modules, lessons, scripts. A production pipeline that front-loads council gates, because a course is a big bandwidth bet.
Command
Run the full gauntlet on an idea, venture, job listing, or offer. Three rounds, the whole council, a 1–5 verdict, logged. Use for any real "should I do this?" decision.
Command
The sabotage-pattern check. Fast pass on an impulse or idea, is this a brave values-walk or the user's named pattern firing (e.g. the scatter, a comparison-triggered uproot)? The prime-directive command.
Command
Grade a past decision or gauntlet verdict against what actually happened, and write it back to the log so the historian learns from truth, not just its own past guesses.
Command
Talk to the mirror, you, talking to you. Not a council, not a verdict. Bring a feeling, a restlessness, a fear, or nothing in particular; the mirror reflects you back in your own words and digs for the real motive.
Command
Prep or run a negotiation, job, salary, review, bonus, contract, client, angry customer. Sets the target, the walkaway, the non-negotiables, and pre-arms against the fold.
Command
Put a pull request on trial. A three-round adversarial gauntlet, map the diff, attack it from every angle a real review would (correctness to spec, quality, maintainability, standards, security, scope), with the author defending, then a 1–5 merge verdict. Built for AI-generated PRs that are too big for one human to…
Command
The setup interview. Builds your dossier — the truth the council fights for — in 20–30 minutes. Must be completed before any other champion command will run. Re-run it any time to revise a section.
Command
Run the tape forward. When something can't cleanly align to the north star, or the user pushes back on a verdict, don't reject, simulate plausible outcomes with honest odds and the reasoning behind them, then hand back the constructive move.
Command
War-game an offer negotiation. The champion builds the opposition from research + the email thread, spars them across rounds, then hands you the exact next move to send, in your voice, holding the line you'd fold on.
Command
Read an email and think through what's really going on, subtext, the ask, the power dynamic, and where the fold-risk hides. No council. Optional reply draft in voice.
Command
Fine-tune a marketing effort against the flywheel playbook. A one-round critique panel, not a full gauntlet. Returns concrete fixes (and the rewrite if it's copy).
Command
Build a YouTube video through a three-round gauntlet into a full production package, Round 1 finds the sharpest angle and hook, Round 2 architects the retention structure and writes the script, Round 3 scouts and prototypes the demo and builds the scene-by-scene direction sheet. Output: a scene & direction sheet…
Command
The writing gauntlet. Anything in the user's voice goes through draft, voice-check, and a de-AI scrub, then is handed to the user in markdown for their edit and approval. Rendering or sending happens only after they approve. All writing is theirs, checked by them.
Agent Claude Code
Implements a single PLAN.md task in TypeScript (Next.js or Bun). Builds and self-tests; never commits unless told. Dispatched by /build-loop.