agree-acceptanceAgrees what done looks like before work starts: concrete checks, who accepts, what the agent must not do, plus examples (one failing) for variable output. Use when starting any task whose result someone else will rely on.
baselineRecords how things are today before a change: the measures that matter, each value's source and date, and \"not measured\" where no data exists. Use when starting a change later judged better or worse, or before setting a target.
capture-the-problemTurns rough notes, a complaint, an idea or a request into a problem statement: who is affected, the cost, what is unknown and the cheapest first test. Use when a request describes a solution rather than a problem.
cross-agent-reviewPairs a lead agent with an independent reviewer (a different person, a different model, or the same model in a fresh session), both given the same acceptance card and evidence. Use when work is medium or high risk or leaves the team. The owner decides.
decision-recordWrites a short record of a decision: context, options, decision, who decided, evidence, what would change it and review date. Use when a supplier, design, policy or scale-pivot-stop call needs recording, or someone asks \"why did we do this?\"
delivery-metricsMeasures the four DORA delivery keys from existing git history, deploy logs and incident notes, plus eval pass rate, cost per accepted task and edit rate for AI features. Use when asking how delivery is going. Not for ranking people.
enablement-planPlans how a team becomes self-sufficient with a new process or agent-assisted way of working: capability matrix, pairing, train-the-trainer, and an exit test of one cycle without the builders. Use when results still depend on the people who built them.
eval-setBuilds a small eval set for AI output: 5 to 20 real examples, what each answer must and must not contain, one failing case, how to grade. Use when relying on AI-written text or classifications, or changing their prompt.
evidence-handoffHands work to a person, team, agent or later session with state, evidence, open decisions and next step intact. Use when a receiving owner needs to take over work with its evidence and open decisions. Final task claims use report-with-evidence; Track results use evidence-pack.
evidence-packBuilds the Track evidence pack with sourced results, the Scope baseline and measurement gaps. Use after a pilot or when someone claims an improvement without numbers. A named owner decides whether to proceed to Own, pivot or stop.
example-mappingRuns an example-mapping session: story, rules, concrete examples, open questions, then settled examples as Given/When/Then. Use when building anything with business rules, or when a request is too vague to turn into checks.
experiment-planPlans the smallest test of a hypothesis: method, sample, duration, cost, success and stop criteria, and who decides. Use when a hypothesis card exists and before building, buying or rolling out the full solution to find out whether it works.
graduation-ladderDecides how much autonomy a repeated task class has earned from logged acceptance outcomes: draft only, automatic with notice, or silent. Use when reviewing a task class or considering less supervision. Never loosens send, spend, delete or publish limits.
hypothesis-cardTurns an idea into a testable bet: \"We believe X will result in Y, and we will know because Z changes\", with measure, baseline, target and stop result. Use when a benefit is assumed, or someone says \"this will help\".
label-the-proofLabels each part of a report as literal, inferred or illustrative, and edits the one canonical file in place, never a new copy, since version history keeps the old one. Use when a reader could mistake an example for a result.
learning-loopTurns mistakes and wins into improvements: observe, diagnose, propose, test, promote, check. A rule needs evidence, a test, a contrary case and authority; one case is weak evidence, so name where the rule would be wrong. Use before adding any rule.
maturity-checkPlaces a person on Yegge's eight AI-development levels and a team on n8n's five maturity levels, scoring evidence separately from self-report, then names the SHOTos stage and skills to start with. Use when asked "where am I with AI" or before granting autonomy.
pilot-before-scalePilots three or fewer items before a run of more than three, then completes a preflight checklist before the full run. Use for bulk edits, mass drafts, clean-ups, migrations or any run touching many records.
progress-wrap-upWrites the weekly progress wrap-up on an evidence ledger, classing each claimed win as live, gated, merged only or unverified, and ending with asks. Use when reporting weekly progress to a sponsor or team. Final results go to evidence-pack.
report-with-evidenceEnds a task that changed something with a short report and Claims block: what was done, how to check it, what was only planned or unchecked. Use when saying something is done, fixed, sent, deployed or passing.
research-and-assembleResearches from sources, keeps each fact with its source, separates fact, inference and guess, and checks every claim before handover. Use when building briefings, comparisons, reports or proposals from outside material that others will rely on.
resume-packWrites a resume pack when a session ends or context fills: handoff note, exact next prompt, pending manual steps and learnings, with no secrets or raw conversation. Use when pausing unfinished work for another session. Use evidence-handoff when a new owner needs to take over.
retrospectiveRuns a short blameless retrospective: what we expected, what happened, why, and what we change, with an owner and date each. Use when a pilot, release, incident or project phase ends, or after a surprise. Not for appraising people.
rollout-planPlans taking something that worked to more teams: a first wave of three or fewer, what is reused, what each team re-baselines, governance. Use when a scale decision is made, even if asked to go everywhere at once. A draft until the sponsor agrees.
route-the-workWrites a one-line route (asset, job, risk, reader, route, proof) and chooses who does the work: direct, a skill, a role agent or a reviewer. Use when starting any task others will rely on.
runbookWrites a runbook from what was actually done: owner, steps, checks, failure handling and rollback, so a team that did not build it can run it. Use when a process becomes someone else's job or depends on one person's memory.
skill-quality-gateRuns a quality gate before promoting a skill or change: one change, one comparison, three scenarios, a maturity label and a disclosure checklist. Use when changing, promoting or publishing any skill, prompt, template or agent instruction.
stakeholder-mapMaps who is affected by a piece of work, who decides and who can block it, with goal, influence, stance and labelled links. Use when starting a change, project or deal. Roles only, not private profiles.