ntholm86

hansei

Reflect on the improvement loop itself. Examine the trail, surface recurring findings, blind spots, methodology effectiveness, and what the loop has been ignoring. Meta-improvement - improving the improvement process. USE WHEN: reflect, retrospective, what are we missing, blind spots in the loop, why does this keep recurring, hansei, examine the trail, meta-improvement, loop critique, what hasnt worked, why is the score plateauing.

intent

Apply Commander's Intent to the user's own prompt before acting. Interpret what the user is trying to achieve, not what they literally wrote. Narrate the interpretation so the user can correct drift before work begins. USE WHEN: any substantive request that implies work (build, fix, improve, explain, investigate, decide). SKIP WHEN: the request is unambiguous and mechanical (a specific file read, a one-line command, a yes/no confirmation).

kaikaku

Radical redesign evaluation. When incremental improvement has converged too low or the architecture is fundamentally wrong, Kaikaku determines whether to redesign and produces a migration plan if warranted. USE WHEN: redesign, start over, rethink, rewrite, architecture is wrong, kaizen isnt working, converged too low, fundamental change, kaikaku, radical change, clean slate, rearchitect, pivot.

kaizen

Incremental improvement - diagnose, challenge blind spots, prioritize by impact, implement, verify. The core improvement cycle. Includes diagnostic vocabulary for unevenness, overburden, and waste. USE WHEN: audit, review, rate, improve, make impressive, quality loop, iterate, kaizen, evolve this, what would make this better, ROI analysis, blind spots, what am I missing.

kata

Orchestrate an improvement cycle: diagnose, select methodology, execute, record, persist. The meta-pattern that connects all skills into a coherent workflow. USE WHEN: improve, audit, review, full treatment, kata, run the loop, comprehensive improvement, what does this project need.

kiroku

Evidence trail management. Start sessions, record decisions during work, close sessions, index decisions, validate trail integrity. The implementation of Observable Autonomy (Principle 2). USE WHEN: start session, record trail, kiroku, audit trail, evidence, close session, validate trail, begin work, observable autonomy, track decisions.

shiken

Construct novelty probes that distinguish genuine situated reasoning from pattern-matching. Builds examination scenarios where routine execution fails but interpretation succeeds. Measures Autonomous Reasoning Fidelity. USE WHEN: test reasoning quality, construct novelty probes, is the agent actually reasoning, ARF measurement, shiken, novelty injection, anti-compliance test, distinguish reasoning from pattern-matching, stress test.

improve

The improvement skill. Understand the ask, examine the target, challenge the first read, decide on one change (or argue for redesign, or declare silence), act, reflect on the target, and record. Combines incremental refinement, structural rethinking, and reflection on the target itself. USE WHEN: improve, audit, review, fix, refactor, redesign, evaluate, what would make this better, am I missing something.

probe

Construct a novelty probe that distinguishes genuine situated reasoning from pattern-matching against a checklist. Build a pair of cases that look similar on the surface but differ in a material way; observe whether the agent's response diverges where it should. Measures Autonomous Reasoning Fidelity. USE WHEN: test reasoning quality, is the agent actually reasoning, distinguish reasoning from compliance, stress test, novelty injection, ARF measurement.

retrospect

Read the trail as a single document and form arc-level claims about the target. What is the target becoming? Where has the loop's attention been, and is that where the target's real weight lies? What does the arc reveal that no individual iteration would surface? Writes .trail/retrospect.md — the Retrospect-derived current orientation for the target. Destination (.trail/destination.md, with .trail/vision.md as legacy fallback), if present, is the operator-held destination and is read but never written. USE WHEN: about to declare convergence, recurring finding-class suspected, operator asks "how are we doing?", or an independent arc-read is needed without running a full improve loop.

trail

Evidence trail management. Append a structured entry to .trail/audit-trail.md IN THE TARGET REPO ROOT at the end of every substantive session — recording the interpretation of the ask, examination, decisions, actions, and reflection. The implementation of Observable Autonomy — autonomy without evidence is not delegation, it is abdication. USE WHEN: any session that produces a decision, realization, or finding — including conversations. There is no such thing as "just conversation" if a decision was made in it.

vision

Surface the agent's in-progress guesses about where the operator is heading — what they care about, what they are circling, what the implicit destination might be — and turn those guesses into questions the operator can confirm, correct, or reject. Closes the gap between what the operator has explicitly stated (vision) and what the agent has picked up from their conversation, reactions, and emphasis. USE WHEN: vision feels thin or stale, the operator is exploring rather than executing, the agent suspects it is missing implicit direction, or before a long autonomous run that will drift if the destination is unclear.

de-ai

Strip the language patterns that mark prose as machine-generated, so the reader engages with the content rather than the style. A diagnostic lens, not a banned-words list. USE WHEN: polishing public-facing prose, reviewing AI-drafted text, applying a finishing pass after Improve has settled the substance. SKIP WHEN: editing verbatim records (transcripts, trail entries, quoted material), or when the prose has already been edited by a human and the AI-tells are operator voice.

destination

Surface the agent's in-progress guesses about where the operator is heading — what they care about, what they are circling, what the implicit destination might be — and turn those guesses into questions the operator can confirm, correct, or reject. Closes the gap between the destination the operator has explicitly stated and what the agent has picked up from their conversation, reactions, and emphasis. USE WHEN: the destination feels thin or stale, the operator is exploring rather than executing, the agent suspects it is missing implicit direction, or before a long autonomous run that will drift if the destination is unclear.