Skills

This page is generated from the plugin SKILL.md frontmatter at build time — the shipped skill descriptions are the source of truth, so the roster here cannot drift from the installed plugin.

The Claude Code plugin ships seven skills in two families. /setup and /interview are the human-driven front doors; the rest are both-invocable — usable from the / menu and auto-invocable by the model. Fidelity measurement is /sync --verify; the on-demand full stock-take is /sync --inventory. There is no command for everyday graph work: once a workspace exists and the session has started with the MCP server wired, you just talk to Claude and the memstead_* MCP tools stay live. A session that is already running picks the new skills up only after /reload-plugins or a restart. Restart the agent session afterwards: a session that is already running does not attach an MCP server added while it runs.

Onboarding & context

Getting a workspace started and its knowledge in.

/setup

Human-only (front door)

First-time Memstead setup in the current directory — resolves the binaries (installing if needed), creates the workspace and mem, and wires the MCP server. Run once per project, then restart the session so the server attaches. Afterwards there is no command for everyday graph work — just talk to Claude; the MCP tools need no slash command.

/interview

Human-only (front door)

Conversational knowledge capture — you talk, Claude asks one question at a time and writes what you know into typed, linked entities. Use when the knowledge lives in a person’s head rather than in any source Claude could read.

/learn

Both-invocable

Load what a mem already knows into context before starting work — searches entities by topic and reads them fully, so the session builds on captured knowledge instead of re-deriving it. Not a docs or web lookup: it reads your mems.

The mem lifecycle

Building a mem from sources, then keeping it true.

/ingest

Both-invocable

Build a mem from a declared source binding — one focused batch per run, resumable, suited to running on a loop. No binding yet? A short three-question setup creates one; no configuration vocabulary needed. Build only: it grows the mem — keeping it current as the source changes is a separate job.

/sync

Both-invocable

Your source changed — bring the mem up to date. Reads what changed since the last run plus any open findings, and updates only the affected entities, conservatively. The single maintenance writer for bound mems; run --all on a loop to keep every bound mem current (the loop ends itself once every binding reports quiescence), --verify <binding> for a fidelity report (coverage, accuracy, freshness) that writes no entity but does record findings and a verified baseline, or --inventory <binding> for the on-demand full stock-take — measure the whole binding, repair to quiescence, report — or --sweep <mem> [<mem>...] for the standing-claim walk: verify what the mem asserts even where no source-change signal points, entity by entity, section by section, one mem after another in the order named. Not a version-control operation: changes flow from your source into your mem, never the reverse.

/remodel

Both-invocable

Restore a mem’s model truth — is the mem still cut right for its subject: does every obligation have exactly one home entity of the right type, does substance sit in its declared sections, is the graph wired. Where /sync repairs what entities SAY, /remodel repairs what the mem IS. Run --all on a loop like /sync: a cheap signal scan walks every mem and descends into the expensive round only where signals justify it (quiescent otherwise — the loop ends itself). Target one mem with remodel <mem> [<cluster>], or measure without writing via --scan [<mem>]. The round derives a target inventory from contract plus source, has it adversarially checked, diffs it against the live entities, rebuilds conservatively, and brackets big rebuilds with a before/after reconstruction probe. Rare and deliberate by design — the refactoring beside /sync’s bugfixing.

/tidy

Both-invocable

Graph hygiene for your mems — assesses structure (orphans, stubs, thin or missing links, stale entries), proposes concrete improvements, and applies only what you approve. Works entirely through the graph: it never reads or writes your sources.