AI Agent Workflow Collection

Designs and audits the system around an autonomous or scheduled agent — everything but the prompt text — and emits it in the target's native form. Trigger to design, spec, harden, review, or audit an agent, bot, scheduled task, or automation acting on its own; to write a Cowork task, a Claude Code routine, or a desktop scheduled task, or the same on ChatGPT, Gemini, or a workflow runner; for guardrails, kill switches, cadence, retries, failure handling, protected resources, output contracts, or handoffs; to security-scan an agent's tool grants, credentials, or blast radius; when untrusted content — email, web pages, documents — needs isolation in an agent; or say agentwright (subcommands emit, audit, security-scan, refresh). Prompt text is promptwright's; standing config a human reads in session — Project instructions, CLAUDE.md — is rigwright's; skill packages as built are skillwright's; code-level threats belong to a security harness.

Defines a brand and its voice — identity, naming, palette, taglines, firewall — and applies them on request across skills, packs, artifacts, docs, and repos. Ships neutral — no brand exists until one is built or handed in; outputs spec-clean. Trigger to create, define, rebuild, or consolidate a brand, voice, or style guide; to apply a brand or voice to a built skill, artifact, repo, or doc; to audit a repo, tree, or skill set for drift — wrong names, off-palette colors, off-voice copy, stale handles or taglines; to export a voice profile or HTML brand-guide card; or say brandwright build / apply / audit / export. Applying a voice to one message is commwright's via the exported profile; a whole skill set's rebrand is a handoff — brandwright defines the identity, skillwright port propagates it.

Shapes any message to its channel and audience. Drafts human by default in a neutral professional voice; a specific brand voice applies only when named or handed in — voice definitions live in brandwright, and defining or saving one, even as "commwright voice", routes there. Trigger to write, rewrite, reshape, or shorten an email, text, Slack or Teams message, release notes, YouTube title, social post, or Discord announcement; to make it more formal or casual; to humanize a message that reads like AI or ChatGPT, stripping em dashes and emoji — a README, CLAUDE.md, or reference doc is skillwright’s prose, not commwright’s; for release comms or a comms plan; or when they name "commwright" for message or channel work ("commwright formats", "commwright audit", "commwright humanize"). It never sends.

Authors and audits eval suites for skills, prompts, and agent specs — a build-time generator whose suites live in the target and run by hand without it. Trigger when someone wants trigger evals, test cases, an assertion suite, or regression coverage written for a skill, SKILL.md, prompt card, or agent spec; when a should/shouldn't set needs balancing or a suite needs scoring — coverage per entry point, boundary pairs, count integrity; when a suite should be refreshed after its target changed; or when they say evalwright (audit — score a suite, refresh — re-derive after changes). For building the skill itself, skillwright; for the prompt under test, promptwright; for code unit tests and QA, engineering test tooling; automated benchmark loops belong to skill-creator's eval tools.

Turns research into verified knowledge products, two modes. Trigger when someone wants a researched recommendation, comparison, or go/no-go — which X should I pick, is Y worth it, compare A vs B — sources checked live, every claim evidence-graded; when they want a reference doc, guide, or playbook, template-first, versioned against primary sources; when a doc needs verification or overlapping docs need consolidating; or when they say lorewright (lorewright verdict / lorewright playbook pick the mode). Verdict ends in one direct recommendation, never a hedge. For prompts, promptwright; for skills, skillwright; for shaping a message to a channel, commwright; broad multi-source research reports are a research tool's job — lorewright produces decisions and reference docs, not reports.

Builds, scores, hardens, and red-teams LLM prompts — from a rough idea to a copy-paste-ready artifact — and picks which model tier to run a prompt or task on. Trigger to write, fix, improve, debug, red-team, or rewrite a prompt, meta-prompt, template, or system prompt; to assemble task parameters into a working prompt; for agent or bot instructions; when asked which model or tier a prompt, a live task, or each subtask of a plan should run on — a plan gets a per-subtask target table that also binds subtasks added mid-session; or say promptwright (promptwright model for a standalone tier and model pick or a plan's target table, promptwright refresh to update model data). For building or auditing skill packages rather than prompts, skillwright; for pure token or cost trims that keep behavior unchanged, tokenwright; a sourced multi-model product comparison is lorewright's verdict, not a run-target pick.

Builds the standing configuration Claude reads before work — a Claude Project's instructions and knowledge-file plan, a CLAUDE.md, a repo's .claude layout, hooks, and .mcp.json — emitted paste-ready in each surface's native form and checked against that surface's real limits. Trigger to set up, write, fix, or trim a Claude Project, project instructions, a CLAUDE.md, or a repo's Claude config; to decide which layer a rule belongs in when the choice is between profile preferences, project instructions, CLAUDE.md, a skill, a hook, or auto-memory; or to score an existing setup for bloat and drift without rewriting it. Answers to "rigwright", "rigwright audit", and "rigwright refresh". For an Agent Skill or a SKILL.md package, skillwright; for anything that runs unattended — a Cowork task, a routine, a scheduled task, and the cadence and guardrails around it — agentwright; for the wording of an instruction block once its home is settled, promptwright.

Builds, audits, ports, and integrates install-ready Agent Skills passing best practices. Trigger to build, audit, score, or package a skill or SKILL.md; to design a pack; when asked if a skill fills a real niche; for a prose pass on a skill's or pack's own files (README, CLAUDE.md); when a skill set needs porting, renaming, rebranding, or sanitizing for a new owner; when a member change must propagate across a pack; or on skillwright (refresh, port, pack, integrate, upkeep). Every build ships trigger evals. Audit covers security — injection surface, secrets, undeclared tools, unsafe defaults. For prompts not skills, promptwright; to define, apply, or audit a brand or voice, brandwright; audits cover the skill package as built — what an autonomous agent may do at runtime is agentwright's.

Measures and shrinks the token footprint of LLM-facing artifacts — prompts, SKILL.md bodies, agent specs, CLAUDE.md, reference docs — at build time, cutting cost without changing behavior. Trigger to slim, compress, or token-optimize an artifact; to fit a token or context budget or a cached prefix; to ask why a prompt or instruction file costs so many tokens; when a set needs budgets or a load plan ("tokenwright budget"); to score token waste without rewriting ("tokenwright audit"); or when they say "tokenwright" ("tokenwright refresh" re-verifies ratios and cache mechanics). Covers a waste taxonomy, a lossless/lossy ladder, and net-cost accounting. For prompt quality or wording, promptwright; for skill best-practice conformance, skillwright; for shortening human-facing messages, commwright; a config's layer placement or setup audit is rigwright's — tokenwright only cuts its cost.

Alex's claude.ai window into Project Longshot (the private MickMacPW/longshot repo) — read today's Daily Bet Card and explain its picks, show bankroll/ledger status and the Monday dashboard numbers, record what Alex actually bet (the placed column), and capture coaching notes that train the model. Trigger on "today's card", "what's the card say", "cardcaller", "log my bet", "I placed/skipped a bet", "how's the bankroll", "coach the model", or dashboard/results questions about Longshot. It never runs the daily pipeline (the cloud routine "Project Longshot - Daily Card" owns that), never places or automates bets on any sportsbook, and never invents a number — if the repo can't be read, it says so and asks for a paste. Not for building betting models or general sports chat.

Runs one pass of the Project Longshot daily NFL bet-card pipeline in the longshot repo — reconcile yesterday's results, update ratings and the ledger, fetch today's slate, FanDuel lines, injuries, and playing-time news, produce the Daily Bet Card, and commit. Trigger on "daily bet card", "today's bets", "linecaller", "run linecaller", or the scheduled daily run. Decision support only — it never places bets, never touches a sportsbook account, and never invents a number; missing data means PASS. Not for building betting models (the repo's code owns that), general sports chat, or work outside the longshot repo.