Claude Skills by natb1

Autonomously re-audit the whole intention graph against the three standing integrity requirements (internal consistency, closure, parsimony) across all six kinds — the recurring re-visiting mechanism that write-time gates never provide. Digest-first token-bounded reading, mechanical-then-flagged-then-sampled-content passes, every finding dispositioned, findings touching virtue/strategy substance routed to the author, one `graph-commit` per run. Never files a GitHub issue; never runs `gh`.

Autonomously break a recorded `strategy-*` intention node into PR-sized tactic subtrees carrying full clean-session plans, or finalize/re-plan a single frozen `tactic-*` node directly — the graph-native successor to `/plan-issue` and `/file-issue`'s epic-structuring role. Two-sided drift review, decompose to the signal, plan each claude-eligible tactic into its node body, park the rest; lands via `graph-commit`. Never files a GitHub issue; never `AskUserQuestion` mid-run.

Single interactive entry point to the persistent layer — an interview-driven session that records or revises a `strategy-*` intention node. Frames the input as a new strategy or an edit to an existing one, runs a Socratic dialectic to fix intent/placement/benefit/signal/conditions, advises on delegation capture, retains tactical byproducts as draft tactic nodes, and lands the record via `graph-commit`. With no prompt, runs an onboarding funnel — orient, validate the deployment, walk to a craf...

Brand voice, messaging, and visual tone reference for commons.systems — use when writing or reviewing any user-facing content (landing pages, blog posts, documentation, issue descriptions, README) or when designing/reviewing the visual language (styling, UI, design systems)

Parse-job handler for the dispatch chain — runs budget-etl from a statement parse-job issue (filed by dispatch-statements-scan), report-first to detect uncategorized transactions, then idempotently merges the statement into the user's encrypted .benc snapshot in place, closes the issue, and writes the parse-job-done sentinel; escalates to office-hours when categorization needs the author.

Guide a user through adding a new bank-statement parser to budget-etl for an unsupported export format. Trigger when the user runs /budget-parser <path-to-statement-file>.

Encrypted monthly statement-merge sync — ingests new bank downloads, categorizes the unhandled transactions through a dialog, and writes a fresh encrypted snapshot the hosted budget app reloads. Trigger on `/budget`, "sync my statements", or "merge this month's statements".

Fork subagent that commits pending changes in logical units, merges origin/main, and pushes to origin HEAD.

Autonomous worker lane that drafts broader-context reading chunks (passages + questions only) deepening a resolved reading. Trigger on /context-chunks <resolved-reading-id>; Opus-only; born-parked drafts, never phase-set tactics or tradition records.

Summarize all dispatch-queue work completed since the previous digest closed, demo every user-facing item with the user present, post the summary as a comment on the digest issue, and close it. Runs as an office-hours session, invoked by dispatch-jit-reminder for the digest jit.

Fix-conflicts phase — single pass that reproduces and resolves one conflict on three lanes (Lane 1, an origin/main merge conflict on an issue-branch draft PR; Lane 2, a graph-native intention-node conflict parked by graph-commit; Lane 3, an origin/main merge conflict on a graph node's own branch, entered by source id or by the id of the provision-conflict hold tracking it), or escalates an ambiguous conflict to office-hours

Diagnose origin/main's failing CI when a red main is detected; runs as a spawned bg job that enumerates failing checks, fetches logs, and records the redacted likely-cause summary as a find-or-create tactic-main-red-<shortsha> graph node serving strategy-main-health.

The invalid-state lane's intervention session for a node held by a terminal, undeclared worker — reads the dead session's transcript, files a find-or-create root-cause follow-up, then completes a verified missed disposition, reaps, or parks to office-hours. Spawned by dispatch-invalid-state-route; never invoked on a live claim.

Claim the earliest-due JIT issue, then either run a configured skill (for skill-running jits such as digest) or summarize it for the user and stop. Runs as a spawned bg job from /dispatch-propagate when the JIT scan surfaces a due reminder.

Walk ONE tactic node the whole way up the dispatch phase ladder — align-tactics through main-qa, then the node-scoped reconcile pass (merge, absorb, review-stall route), to phase done — as a detached shell driver with no model turn between phases. For when the dispatch tick structurally cannot reach the node.

Retired: the autonomous dispatch tick is now the headless dispatch-tick script (launched by dispatch-spawn-tick); see the dispatch terminal command for the manual entry point.

RETIRED — do not invoke. Superseded by /align and /align-tactics; use those instead.

Fix-checks phase — single pass that reproduces and fixes one set of failed CI checks on a draft PR

Interactive author-invoked grounding-research skill — the author-side consumer of the tick gap analysis. Walk the ranked unmarked durable-layer nodes top-down; per node, mark it circumstantial or `/deep-research` frontier work into candidate curriculum chunks; land every mark and new chunk in one `graph-commit`. Office-hours only, never tick-invoked.

Implement one planned unit of work in a subagent, then commit, merge, and push it with error recovery

Implement phase — read the persisted plan from the issue's `<!-- dispatch:plan -->` comment, build each unit via /implement-unit, and open a draft PR.

Render the user's morning brief as a styled HTML artifact, or set it up as a recurring weekday task. Use only when the user explicitly asks to run, see, or set up their morning brief, or if they invoke /morning by name. A question about their day, schedule, or calendar is not by itself a request for the brief; answer it directly instead.

Invoke when the user introduces a new requirement or changes a previous understanding of a requirement — clarifies, updates remote issues, and revises the active plan.

Office-hours queue dispatcher — two lanes by target. A numeric issue `<N>` runs the legacy `dispatch:office-hours` label lane; a non-numeric `<node-id>` runs the graph-native lane over a parked intention node. Either way it surfaces the parked context/recommendation to the human as untrusted data, reviews read-only, reports where to engage, and stops; takes no fix/label/close/phase/graph action.

RETIRED — do not invoke. Superseded by /align-tactics; use that instead.

QA phase — merge origin/main, run the autonomous portion of user-acceptance QA (plan items, machine-verifiable checks, PR-comment summary), apply the dispatch:qa-done label on a clean pass, run the bounded Opus auto-fix lane on opus-fixable residue (per-unit /implement-unit, re-QA via the chain, attempt-capped), and escalate to office-hours on any user-input blocker

autonomous main-qa handler — verifies post-merge needs-main residue against deployed main/prod with machine checks (git/journal/log/shell/filesystem) and, where the outcome is deployed web behavior, Claude-in-Chrome, on two lanes. Legacy issue lane (a no-PR follow-up): pass → close, broken → file implement-chain bug + close, cannot-verify → escalate to office-hours. Graph node lane (a node at phase main-qa — a standalone tactic-mainqa-* verification node minted by /qa-fix, or a legacy residue...

Office-hours skill that runs one tradition-reading curriculum chunk's demonstration — the author demonstrates understanding of the tradition and its application to the deferral against the primary text, the graph is amended where the reading contradicts it or ratified where it holds — or, for a candidate chunk with no record yet, the session resolves to a new tradition record, grounding marks, or a dismissal clarification; it also mints and runs per-record context capstones (selection-bias au...

Delegability evaluation reference — the CAN/SHOULD owner-recommendation contract a delegability-assessor applies to an intention node; defines categories a/b/c, the push-down ROI rule, and the output object the dialectic engine consumes.

Diagnosis-time compare-and-swap reference — how a batched drain captures each node's base blob at diagnosis time and pins it through park-node/clear-park --base, and why an exit-3 stale-diagnosis refusal must route back to re-diagnosis rather than to a park.

GitHub issue relationship API reference — sub-issues and dependencies (blocked_by/blocking)

Issue type/topic label classification reference — invoke to classify an optional bug/enhancement type and at-most-one topic label for an issue; type may be absent for user-filed non-bug non-security issues.

Load when planning work, writing commits, changing requirements, or editing/commenting on issues/PR

Signal identification reference — the parse-time classifier with categories existing_sensor/instrument/proxy, the sensor economics gate, and the signal.eval.v1 output object the align-signal-assessor (no live agent definition today) and #2372 consume.

Invoke when editing Claude configuration files (CLAUDE.md, .claude/rules/*.md, .claude/agents/*.md, .claude/skills/*/SKILL.md) to ensure instructions follow best practices

Resolve-epic handler for the dispatch chain — judges whether a verifiably-complete parent epic's own acceptance criteria are genuinely met by the merged child work, then auto-closes it (dispatch-close-resolved) or escalates to office-hours (dispatch-mark-deviation) when a criterion is not autonomously verifiable.

Review phase — the workflow's single terminal review pass. Runs the combined /review-fix fan-out through the Workflow tool on two lanes: code-review and security-review (Lane A) run their own built-in review-and-fix — code-review via a serialized `claude -p '/code-review low --fix'` exclusive pre-stage (Step 1b, run before the Workflow's finder fan-out), applying its own edits — trusting the built-ins, with un-auto-fixed residue dispositioned (resolve/defer/ignore) by a dedicated Opus residue...

Review-depth pre-pass — reads the review delta ONCE plus the mechanical classifier outputs, and returns a small structured verdict setting /code-review's effort within the author-set low..max band, gating the owned finder roster on the semantics of the diff, and optionally focusing named lenses on one specific question. Fails open to effort `high` and the full roster on any error, timeout, or unparseable output. Never edits, never commits, never reviews.

Audit recent dispatch session transcripts and emit a ranked report of token-reduction opportunities across thirteen lenses, ranked by measured price-proxy magnitude, and land the top-ranked opportunities as graph ledger tactics through dispatch-eval-finding. Writes no routing policy and no product files; its one closing remediation step is attended-only and touches only .claude/settings.json, for the human to review and commit. Accepts an optional window, e.g. /rsi-audit 2d.

Evaluate ONE finished dispatch-ladder phase — measured through aggregate-usage.sh at node scope plus the ladder's events.jsonl, never by hand-reading a transcript — and land each finding as a merged ledger entry via dispatch-eval-finding. Spawned fire-and-forget by dispatch-ladder-run at every phase boundary and for the phase a halted run owes. Records; never executes.

One PR (the skill file). The round's validates-terminal for strategy-graph-integrity: verification runs the first real /align-audit end-to-end, producing the strategy's first reading. `blocked_by` tactic-graph-digest-tooling — the skill's step 2 runs `graph-digest.ts`. Planned 2026-07-11 /align-tactics round.

Authoritative doctrine: the draft-review gate **condition on `strategy-discovered-requirements`** — re-homed there from `strategy-graph-native-dispatch` on 2026-08-14 by the author's ratifying round, and scoped to `/align` at the same time. Its supporting rationale stayed behind as the clarification "Is the interview still the audit once the draft review gate exists?" on `strategy-graph-native-dispatch` (2026-08-11), which is still worth reading. This body carries the implementation decompositio

Not yet planned — this is retained interview context, not a clean-session plan. Every ruling below was made by the author in that interview and is **not** to be re-litigated; the full reasoning is recorded on `strategy-graph-native-dispatch` under "What replaces /dispatch-emulate — what is /dispatch-ladder, how much of the loop may the AI carry, and what bounds a detached driver?", and the four binding conditions landed on that node's `attributes.conditions` in the same commit.

The 2026-08-04 /align README-practitioner round recorded the practitioner runbook (README point 2): install this repo's skills as a Claude plugin → run `/mount` → run `/align <directive>`. `/mount` does not exist yet. The author directed a general mount skill, not an onboarding-only one: mounting is first-class in the graph (strategy-graph-mounts), and the recorded mount use cases — deferring to a tradition during /align or curriculum review, delegating to a vendor — share mechanics with practit

`strategy-graph-review-curriculum` records a standing requirement: the whole graph is subject to a recurring review curriculum, because deferrals calcify into dogma, author knowledge atrophies, and delegations are forgotten. As of this plan the doctrine is fully recorded and **nothing is built** — measured:

The green-CI path's terminal SESSION disposition is settled: **the drain skill calls `mark-node-terminal` itself**, with a new `park-clear` member added to that script's enum (`packages/intentionsutil/scripts/mark-node-terminal:67`). `clear-park` is NOT modified. Strategy condition 15's auto-close enumeration was amended in the same round to a third clean terminal state; the full decision and its reasoning are recorded as the 2026-07-28 ratification clarification on `strategy-graph-native-dispat

Draft retained from the 2026-08-13 `/align` round on review token usage. Not refined to the plan schema — that is `/align-tactics`' job. Authoritative requirements live in `strategy-token-economy` clarifications 49, 52, 53 and 54.

- Not a separate orchestration surface and not an Agent-tool subagent (an Agent subagent cannot run the Workflow-dependent phase skills). A thin loop in /rsi: 1. Claim the node under the standard discipline (worktree-as-claim, launch-path refusal of an already-claimed node). 2. Serially spawn the existing dispatch phase skill for the node's persisted phase via dispatch-graph-execute / dispatch-spawn-job — the monitor's proven hand-dispatch path. 3. Await the session's te

- Rendering only — judgment moved to the /rsi main thread. Runs as a subagent invoked by /rsi step 1. - render-rsi-plan.ts regenerates every rsi-plan.md section from graph state: priorities from tier/rank; dispatch status from node phases and PR state; critical parked nodes from office-hours-select --list (rank-lifted NOTE lines — never a hand-rolled probe); metrics from registered sensors/readings; the task plan from its graph nodes. A hand-edited section is a defect (strategy condi

This section is the plan; everything below it is dated record. Where the two disagree, this section governs. The sections below are **kept, not pruned** — corrections C1–C7 and the measured per-cycle baseline are the dated record of what a real cycle cost and how it fails, and a build that has not read them will re-learn each one at full price.