Conciseness

The body never explains concepts Claude already knows and is information-dense, but several sections are design-history justification rather than operational guidance: "It used to be narrated", "Why not attribute automatically? It was tried and it does not work", the Stop-gate disarm rationale, and the dravr-platform/mirroir-mcp sweep comparison. These could be cut or heavily compressed without affecting execution, fitting the 3 anchor (mostly efficient, some unnecessary explanation that could be tightened) rather than 4 (only minor trims needed).

Actionability

Concrete, executable commands appear throughout: the four bilan.sh invocations with flags, defined exit codes, `bilan.sh ack --why "..."`, `git push origin <branch>:main`, `git submodule update --init .registre`, and `carnet.sh status <n>`, plus the precise core rule "Run it before you give a number. Report the score it prints." Minor gaps keep it below 5: most caps' remedies are described in prose rather than given as commands, and the limitation-label procedure names subcommands without a full invocation.

Workflow Clarity

The core loop — run the script, report the number it prints, apply each cap's printed evidence and remedy, re-run — is unambiguous and deterministic (exit codes, the min() rule), and the limitation exemption has an explicit two-half validation. It sits below 5 because the sequence is implicit across prose sections rather than a stated workflow, and the sweep/worktree subsections describe conditions and edge cases rather than ordered steps.

Progressive Disclosure

No bundle files exist (no references/, scripts/, or assets/ directories), and all deep reference material — repo shape, sweep session-forensics, worktree/submodule mechanics, CI policy — is inline in one ~310-line file. Section headers are clear, so it is above the 2 anchor's minimal structure, but content that belongs in separate reference files is inlined and there are no one-level-deep references at all, matching the 3 anchor.

Specificity

"Measure whether this session actually finished — the 0-10 completion number, computed from git, the carnet ledger, LIMITATION markers and CI instead of narrated" plus "find what a session that died left behind" lists several concrete capabilities with their evidence sources. It falls short of the 5 anchor only because it does not enumerate the concrete checks (unpushed commits, todos, background tasks) that the body shows; it clearly exceeds the 1-2-action coverage of the 3 anchor.

Completeness

Both halves are explicit: the what ("the 0-10 completion number, computed from git, the carnet ledger, LIMITATION markers and CI instead of narrated") and the when with concrete trigger phrases ("Use before reporting a completion number, when asked 'where are we', at the end of a piece of work, and to find what a session that died left behind"). This matches the 5 anchor — the 'when' clause is explicit and multi-trigger, not the merely-present 'when' of the 4 anchor.

Trigger Term Quality

Natural user phrases are present: "reporting a completion number", "where are we", "at the end of a piece of work", "find what a session that died left behind". A few natural variations users would say are missing ("are we done", "did it finish", "wrap up"), which fits the 4 anchor (good coverage, a few terms missing) rather than the comprehensive 5.

Distinctiveness Conflict Risk

The niche is highly specific — session-completion scoring keyed to named internal systems ("carnet ledger", "LIMITATION markers") and a "0-10 completion number" — and every trigger is anchored to that number, so it is unlikely to fire for unrelated skills. Clear niche with distinct triggers, minimal conflict risk.