Conciseness

The body restates the same invariants many times over — "an index's gist is a cue and supports nothing" appears in the MORPHISM block, the ClaimSupported axiom, the grounding table, and the Source-grounded recognition rule, and "knowing it took place, not that it still holds" is repeated across the FLOW, RecalledContext, presentation, and rules sections. The 535-line Lean contract plus four prose restatements of the same flow (FLOW, MORPHISM, PHASE TRANSITIONS, LOOP) is heavily padded relative to what an executing model needs, though it does not explain concepts Claude already knows.

Actionability

There is concrete guidance — the condition-triggered reference-loading sequence ("read its realization reference (references/claude.md or references/codex.md)"), a failure-mode taxonomy with recognizable triggers ("FalseAnchor — an identifier matched, and its record is not the one meant"), and an explicit presentation checklist — but the operational core is a Lean specification whose axioms are declared "black boxes to the contract, yours to make", and the body contains no executable commands or worked example of a round. This is pseudocode-level guidance with key execution details deferred, matching the anchor for some concrete but incomplete guidance.

Workflow Clarity

The FLOW block gives an explicit branching sequence (pass reads the cue, runs searches, opens the leading recognizable, then presents / asks / closes unresolved) and classifies every subsequent utterance (identified / Elsewhere / Reaches / session answers), with checkpoints for scope reports, dissent, and capture qualification. It falls short of anchor 5 because error recovery is deferred to references and the actual sequence must be assembled from four separate restatements rather than one explicit validated workflow.

Progressive Disclosure

All seven reference files referenced in the body exist in references/ and each is loaded under an explicit condition ("read references/supra-session.md before joining records into it", "read references/fork-resume.md before emitting its handle"), one level deep, with scripts reachable via capture-outcome.md. However SKILL.md itself is a ~43KB monolith carrying the full Lean contract inline rather than a concise overview, which is more than a minor organization gap and keeps it below the clear-overview anchor.

Specificity

It names the domain (vague recall of past work) and two actions — "find it in past records" and "to recognize in a story" — but neither is made concrete: what records (conversations, artifacts, decision records) or what presenting a story operationally involves is left unstated. This matches the anchor for one or two concrete actions without comprehensive coverage, not the several-enumerated-actions anchor above it.

Completeness

Both parts are present: the "when" is an explicit trigger situation ("The user vaguely recalls a past discussion, line of work, topic, or settled concept but cannot name it") and the "what" is "find it in past records to recognize in a story, knowing it took place, not that it still holds". It is not anchor 5 because the "what" is phrased obscurely and the capability must be decoded rather than read; not anchor 3 because the when-clause is explicit rather than weakly implied.

Trigger Term Quality

"vaguely recalls", "past discussion", "line of work", "topic", and "cannot name it" are natural trigger terms, but common variations a user would actually say — "remember", "we talked about", "earlier conversation", "previous session" — are missing. This is some relevant keywords with missing synonyms, not good coverage with a few gaps.

Distinctiveness Conflict Risk

The vague-recall-without-a-name trigger carves a clear niche that few skills would fire on, but "find it in past records" is broad enough to carry minor overlap risk with general session-history search or memory skills. This fits mostly-distinct-with-minor-overlap rather than the minimal-conflict anchor.