Operations: planning & review
Operations: planning & review/create-plan [describe what to plan]Operations
What it is. Designs a structured implementation plan for a non-trivial change before any code is written: objective, scope, critical files, step sequence, success criteria, and risks.
What it does. Researches the workspace first (core reference files, existing patterns, connections), then writes a complete plan document to plans/YYYY-MM-DD-{name}.md following a fixed template (Overview, Spec Core, Current State, Proposed Changes, Step-by-Step Tasks, Success Criteria). It plans only and never makes the changes. Steps are grouped into optional parallel waves so /implement can execute them cleanly, and the Spec Core must carry at least one Non-Goal and a single testable Success Signal.
How to use it. Use before significant or multi-file work; to carry out the plan use /implement, and to stress-test it before approval use /scrutinize. Do not use it to execute an existing plan.
/create-plan add a new intake command with routing config
"plan for restructuring the outputs directory"Customize. The one argument is a free-text description of what to plan. No flags. The plan structure is fixed by the reference template; wave grouping is optional inside the plan itself.
/canopus [note | check | probe]Operations 🔒
What it is. The engineering standard as a command: seven numbered steps in four acts, two of which are the operator's own (step 4, his commit of the plan and the red contract, and step 7, his word to ship). It exists to stop one thing, a contract too weak to decide anything passing as one that decided.
What it does. Bare /canopus prints the seven steps and where the operator's two moments sit in them; it runs nothing and reads nothing off the disk, because only steps 4 and 7 leave a trace in this repository and a position inferred from the rest would be a guess dressed as a measurement. The three subcommands are thin. note writes the slice's committed record under records/slices/, and the schema refuses a note missing any required field rather than writing a half-formed one. check runs the four clauses: the contract did not move between the approval sha and the end state, the implementation descends from that approval, the contract was red when checked out at the approval sha and run there, and the target is green at HEAD. probe measures whether a contract's redness means anything, by null-stubbing the modules it imports and running it twice with different stub values: a test that never fails under either run is vacuous. It also puts three wrong implementations that exist in front of the contract and prints what each one took of the red set, and it refuses a skip-family marker that states no reason, whether it sits on a test, on a class, or on pytestmark. probe --after-build <test paths> asks that last question again after the slice has shipped and its contract has been retired: it puts the same three wrong implementations in front of code that ALREADY EXISTS and names every test that stayed green under all three. That page says which modules were actually replaced and which were not, and names a skipped test separately from one that went red, because a test that never ran is neither. It reports and never refuses, and nobody has calibrated it, so it must not be wired into a gate.
How to use it. Run bare first for the agenda; skip the whole standard for typo fixes and config-only edits. The freeze is a commit and the separation is a dispatch, so there is no lock to take and no release to run: step 4 is /scrutinize on the plan followed by the operator's own commit, step 5 is superpowers:subagent-driven-development, step 6 is /scrutinize --relentless, and step 7 is his word plus /canopus note.
/canopus
python scripts/canopus.py probe tests/contract/2026-07-01-add-intake-command/
python scripts/canopus.py probe --after-build tests/test_intake.py
python scripts/canopus.py check --range origin/main..HEADThe honest limit. A step in the sovereignty guards CI job runs the same four clauses on every push. It REPORTS a broken clause; it does not block one, because enforce_admins is false on the only push path in use. Nothing here prevents a contract from being edited by whoever is implementing against it. Do not describe this as prevention.
Customize. Absorbed /pre-impl on 2026-08-02; the planning gate lives in references/planning-gate.md. The thirteen-step freeze lifecycle it carried was retired on 2026-08-07 in favour of git ancestry plus a committed per-slice note, and the agenda now has exactly one definition, scripts/utils/canopus_steps.py. The skill never takes either approval on the operator's behalf.
/implement [plan-path] [--no-trajectory] [--evaluate]Operations
What it is. Executes an approved implementation plan step by step, writing complete files and emitting a structured trajectory for every phase so the run can be audited later.
What it does. Reads the plan in full, runs a soft planning-gate reminder, then executes each step (or each wave, when the plan uses ### Wave headers, dispatching parallel steps as agents). It writes a JSONL trajectory to outputs/operations/implement/_trajectory_<run_id>.jsonl unless --no-trajectory is passed, validates against the plan's checklist, and updates the plan status to Implemented. Pass --evaluate to grade each artifact with /evaluate after execution (up to three rework cycles).
How to use it. Use to build out a plan from /create-plan; do not use it for planning itself. To audit the finished run afterward, run /scrutinize trajectory:<run_id>.
/implement plans/2026-07-01-add-intake-command.md --evaluate
"execute the plan"Customize. The plan-path argument, plus two flags: --no-trajectory (skip JSONL emission for throwaway runs) and --evaluate (run /evaluate on each artifact after Phase 3).
/evaluate [artifact-path] [--plan plan-path]Operations
What it is. A skeptical quality evaluator that grades a workspace artifact (skill, script, reference file, or rule) against workspace standards and, optionally, a plan's success criteria.
What it does. Runs a two-layer assessment: deterministic checks via scripts/artifact-evaluator.py, then a qualitative pass from a deliberately skeptical reviewer persona that looks for shortcuts and vagueness. It ends in one of four grades (PASS, PASS WITH NOTES, NEEDS REWORK, FAIL) with specific, line-referenced rework instructions. A deterministic failure cannot pass.
How to use it. Use to grade an artifact against a fixed rubric after /implement or standalone. For fact-checking a draft against source documents use /validate; for a multi-phase stress-test that proposes and applies fixes use /scrutinize.
/evaluate .claude/skills/intake/SKILL.md --plan plans/2026-07-01-add-intake-command.md
"grade this script"Customize. The artifact-path argument plus an optional --plan <plan-path> to grade against that plan's success criteria and validation checklist. Artifact type is auto-detected from the path.
/scrutinize [plan | execution | file:<path> | dir:<path> | workspace | trajectory:<run_id>] [--relentless] [--no-refute]Operations 🔒
What it is. A maximum-effort principal-engineer review gate that runs a Validate-Identify-Improve-Adjust pass over a target, then presents evidence-backed findings with concrete proposed fixes, blocking forward progress until approved.
What it does. Runs 14 subchecks over the target, then an adversarial refutation layer (Phase 2.5) with cross-family judge rotation (Claude and Kimi k3, with Gemini and Grok reachable via --judge-family) and a two-agent debate on BLOCKER and HIGH findings to drop false positives. Findings carry confidence scores and land in an approval block; only explicitly approved fixes are applied. The workspace target dispatches five parallel specialist agents; --relentless pre-approves fixes and loops with adaptive termination (two-zero, marginal-twice, a 10-iteration cap, check failure, or oscillation), and --no-refute skips the debate layer.
How to use it. Use to stress-test a plan before approval or audit changes after /implement. For artifact grading against a rubric use /evaluate; for fact-checking a draft use /validate; for decision reasoning use /deep-think.
/scrutinize file:scripts/intake.py --relentless
/scrutinize trajectory:2026-07-01_143052_add-intake-commandCustomize. Six target selectors (plan, execution, file:<path>, dir:<path>, workspace, trajectory:<run_id>) plus four flags: --relentless (auto-apply and loop, incompatible with target=plan), --no-refute, --include-low-confidence, --include-ambiguous. Confidence threshold defaults to 75. Frontmatter carries disable-model-invocation: true.
/validate [content or file path]Operations
What it is. Fact-checks a draft against the source-of-truth DataStore documents, bucketing every claim as Validated, Unverified, or Contradicted.
What it does. Extracts every factual claim (numbers, names, dates, technical and market claims, superlatives), then validates each against a source hierarchy where the DataStore wins over context files, which win over reference files. Each verdict carries a forensic evidence grade (Confirmed, Deduced, Hypothesized) and a risk level (High, Medium, Low), and the report ends with corrections prioritized by risk. It never fabricates a validation: no source found means Unverified, not a guess.
How to use it. Use to verify facts, figures, and superlatives before external-facing content ships. For grading an artifact's craft and completeness use /evaluate; for a full engineering review use /scrutinize.
/validate outputs/content/2026-07-01_proposal-draft.md
"fact-check this draft"Customize. One argument: either pasted draft text or a file path. No flags; the source hierarchy and risk table are fixed.
/editorial-review [file:<path> | pasted text]Operations
What it is. A document-level structural editor for long deliverables: it reviews a draft's argument architecture (section ordering, claim-to-evidence linkage, hierarchy, redundancy, buried lede) and never touches the ideas.
What it does. Phase 1 selects a document model, maps the structure with word counts, walks a defect checklist, and emits findings each tagged with one operation (CUT, MERGE, MOVE, CONDENSE, SPLIT, ADD, PROMOTE, DEMOTE, PRESERVE, QUESTION) and a word-savings estimate, behind an approval gate. On approval it applies only the approved structural edits. Phase 2 hands all sentence-level work (rhythm, specificity, vocabulary) to the always-on humanization prose rule and runs the sanitize and humanization-check gates. Content is sacrosanct: it reorganizes, it never rewrites sentences.
How to use it. Use on a long proposal, brief, or report whose argument arc needs checking. For sentence-level "make this sound human" work use the humanization rule, not this skill; for fact-checking use /validate; for artifact grading use /evaluate.
/editorial-review file:outputs/content/2026-07-01_market-brief.md
"review the structure of this draft"Customize. One argument: file:<path> or pasted text. No flags; the operation vocabulary and defect checklist are fixed by the reference file.
/workspace-deep-audit [--mode={full|quick|focus}] [--focus={skills|rules|deps|security|architecture}] [--vs=<previous_audit_path>]Operations 🔒
What it is. The heaviest workspace skill: a comprehensive 8-section deep audit of the entire workspace (executive summary, inventory matrix, architectural findings, dependency validation, a best-practices gap score, competitive mapping, prioritized recommendations).
What it does. Dispatches up to seven parallel read-only inventory agents (batched under the concurrency cap), validates every dependency pin against current docs via Context7, scores the workspace against a 34-point best-practices rubric, maps it against a competitive baseline, and emits prioritized P0 to P3 recommendations. Output renders to markdown plus branded HTML and logs to an audit thread. It is read-only against the workspace: fix application belongs to /scrutinize. --mode=quick skips competitive and best-practices analysis (roughly half the runtime), --mode=focus plus --focus scopes to one subsystem, and --vs produces a before/after delta report.
How to use it. Use at a major milestone or when checking ecosystem drift. For a single-skill review use /evaluate; to apply fixes use /scrutinize; for a quick operational health check use /state-check.
/workspace-deep-audit --mode=focus --focus=security
/workspace-deep-audit --vs=outputs/operations/workspace/2026-05-14_audit_workspace-deep-overview.mdCustomize. Three flags: --mode (full default, quick, focus), --focus (one of skills, rules, deps, security, architecture, observability; required with focus mode), and --vs=<prev_audit_path> for delta mode. Frontmatter carries disable-model-invocation: true.
/calibrate [light]Operations 🔒
What it is. A reflective end-of-session self-improvement pass: it scans the current session for corrections, preferences, repeated patterns, errors, successes, and voice violations, then proposes concrete patches.
What it does. Runs six-category detection over a parsed session envelope, applies an idempotency check, and resolves each candidate's classification. CEO-only candidates go to a numbered approval list; corporate-classified candidates route to a separate review queue and are never auto-applied. Approved patches to memory, settings, and CEO-only skills and rules are applied and committed in one atomic git commit. /calibrate light scans only the in-context conversation (skipping the repeated-patterns and errors categories) for low-token or quick sweeps.
How to use it. Use at the end of a working session to capture learnings. For cross-session memory consolidation use /dream; for grading a single artifact use /evaluate.
/calibrate light
/calibrateCustomize. One optional argument: light (or --light) selects the lighter in-context mode. No other flags. Frontmatter carries disable-model-invocation: true.
/checkpoint [optional note]Operations 🔒
What it is. Saves a manual session handoff so the session can be resumed later with clean context, without running compact or clearing anything.
What it does. Writes one combined handoff file (objective, acceptance criteria, constraints, decisions, files touched, current state, next steps, continuation prompt) to outputs/operations/handoff-archive/, then updates the .latest/summary.md and .latest/prompt.md pointer files that the SessionStart inject hook reads on resume. It does not run compact, does not clear the session, and does not continue work after writing. It surfaces from the two-tier checkpoint-offer hook at 25 and 30 percent context used.
How to use it. Use before switching tasks, before a risky action, or when the checkpoint-offer hook fires. For a reflective end-of-session capture use /calibrate; for cross-session memory consolidation use /dream.
/checkpoint paused mid-refactor, tests still failing
/checkpointCustomize. One optional free-text note argument, appended as a "User note" section (omitted if empty). No flags. Frontmatter carries disable-model-invocation: true.
/align [N]Operations 🔒
What it is. A manual escalation lever above the always-on prompt-refinement rule: it forces exactly N numbered clarifying questions before any work begins, so scope is locked up front instead of corrected later.
What it does. Parses and validates N, walks back to the latest substantive request, then returns a compact expansion plus N numbered questions, each with lettered options and one committed recommendation, a summary table, and a closer. After the user answers (compactly, for example "1a, 2c, 3 - custom: ..."), it produces the full expanded prompt and waits at an approval gate before executing. Single-shot per invocation.
How to use it. Use before a high-stakes or ambiguous task where a wrong assumption is expensive. For simple requests the always-on prompt-refinement rule already handles clarification, so skip it. For contrarian critique use /devil; for content variants use /burst.
/align 3
/alignCustomize. One optional argument N, default 5, valid range 1 to 10; anything outside the range aborts with a one-line error. No flags. Frontmatter carries disable-model-invocation: true.
/devil [N] [: <inline claim>]Operations 🔒
What it is. A manual sycophancy-breaker: it forces exactly N contrarian critique points against a target position, each from a distinct angle, in committed prose with no fixes and no hedging.
What it does. Targets either an inline claim (/devil 5: <claim>) or the latest substantive decision walked back from conversation. It emits N numbered points, each with a severity tag (BLOCKER, HIGH, MEDIUM, LOW), a bolded 2-to-5-word headline, and one or two sentences that land the cut and stop. Points span distinct angles (correctness, scope, cost, timing, alternatives, second-order effects, stakeholder risk, and more) and sort by severity. An honesty floor stops it early rather than fabricate weak points to hit N. Single-shot: after the pass it exits with no approval gate.
How to use it. Use to pressure-test a recent decision cheaply. For external Gemini, Grok, and Kimi views use /council; for a multi-phase principal-engineer review use /scrutinize.
/devil 5: we should ship the intake command without a test contract
/devil 3Customize. Optional N, default 5, range 1 to 10; optional inline claim after a colon (else it walks back). The force N reply is the honesty-floor escape hatch to surface weak points anyway. Frontmatter carries disable-model-invocation: true.
/burst [N] [: <inline seed>]Operations 🔒
What it is. A manual variation lever: it produces N different versions of the latest content artifact, each attacking a distinct axis, with one mandatory "swing-the-other-way" variant that inverts a defining property.
What it does. Targets either an inline seed (/burst 3: <seed>) or the latest assistant-produced content turn. It generates N-1 spread variants across diverse axes (opener, tone, structure, lens, length, voice, metaphor) plus one swing variant, holding the core message constant and varying only the wrapper, then closes with one recommended pick. A cost-and-time pre-flight gate fires when each variant would run a paid-API call or a multi-step task. Single-shot, and it supports the convergence pattern: pick a variant, run /burst again from there.
How to use it. Use to compare directions, escape a stuck draft, or converge on a final version. For N clarifying questions use /align; for contrarian critique use /devil.
/burst 5: a 150-word post about data sovereignty
/burstCustomize. Optional N, default 3, range 2 to 5; optional inline seed after a colon (else it walks back). The cost gate requires an explicit "go" per invocation for paid or long tasks. Frontmatter carries disable-model-invocation: true.
/brain-audit --sources <paths> [--entity <name>] [--modalities <list>]Operations
What it is. A small post-synthesis audit primitive: given a set of cited source files, it returns a stable three-section footer reporting source freshness, modality coverage, and disagreements.
What it does. Runs three checks over the sources passed in: newest-modified date per source (flagging anything over 90 days as STALE), comms and intel modality coverage for a named entity (email, telegram, osint, crm-log, calendar), and a single LLM call to detect disagreements between sources. It returns a ## Brain audit footer and nothing else. It is a leaf: it never synthesizes, never scans the whole workspace, and never writes state, and every failure mode degrades to a valid footer rather than aborting the caller. It is composed by /meeting-prep, /odin (consult mode), and /deal-strategy, which paste the footer beneath their output.
How to use it. Not a natural-language command. A composing synthesis skill invokes it at the end of its synthesis phase; the CEO may also invoke it directly for an ad-hoc audit.
/brain-audit --sources crm/contacts/example-corp.md,outputs/intel/osint/2026-07-01_osint_example-corp.md --entity "Example Corp"Customize. --sources (required, comma-separated workspace-relative paths, capped at 20 for the contradiction check), --entity (optional, scopes the modality scan), and --modalities (optional subset of the canonical list). Footer shape adapts to single-source, no-entity, degraded, and no-sources cases.