Conciseness

The body is dense, imperative, and assumes Claude's competence throughout — e.g. "Spend disproportionate effort here", "Cut inputs, callers, config, data, and steps one at a time", "Untagged logs survive; tagged logs die". No section explains concepts Claude already knows; rationale lines ("Wrong bug = wrong fix", "Why bother: a minimal repro shrinks the hypothesis space") are minimal and load-bearing, matching anchor 5 rather than anchor 4's 'minor instances that could be trimmed'.

Actionability

Guidance is concrete and executable at every step: a ranked list of 10 loop-construction tactics (failing test, curl script, "run 1000 random inputs", "`git bisect run` it"), a real script path (`scripts/hitl-loop.template.sh`, which exists in the bundle), a concrete log-tag convention (`[DEBUG-a4f2]` with "Cleanup at the end becomes a single grep"), and a falsifiable-hypothesis format string. For an instruction-only skill this is fully specific guidance with copy-paste-ready artifacts, matching anchor 5.

Workflow Clarity

Six phases are clearly sequenced with explicit validation gates and checklists: "No red-capable command, no Phase 2", "Do not proceed until you have reproduced and minimised", a 4-item red-capable checklist, "Watch it fail / Apply the fix / Watch it pass / Re-run the Phase 1 loop", and a Phase 6 cleanup checklist. This is the anchor-5 pattern (explicit validation steps, feedback loops, checklists) with no gaps that would place it at 4.

Progressive Disclosure

The body is a well-sectioned overview whose only bundle dependency — `scripts/hitl-loop.template.sh` — is clearly signaled, one level deep, and verified to exist with content matching its description (step/capture helpers, KEY=VALUE output). All inline content (phases, checklists, construction tactics) is the core instruction and appropriately belongs in SKILL.md; navigation via phase headers is easy, matching anchor 5.

Specificity

"Diagnosis loop for hard bugs and perf regressions" names the domain and a single core action (a structured diagnosis loop), but does not enumerate what the loop actually involves (build a repro loop, minimise, hypothesise, instrument, fix with a regression test). This matches anchor 3 ('names domain and 1-2 concrete actions, but not comprehensive'); it is above anchor 2 because the domain is concrete, and below anchor 4 because only one action is stated.

Completeness

Both 'what' ("Diagnosis loop for hard bugs and perf regressions") and 'when' (explicit "Trigger: diagnose, debug this, something broken/failing/slow") are clearly and explicitly answered with concrete trigger phrases, matching the anchor-5 example pattern. Anchor 4 does not fit because the 'when' clause is already explicit and phrase-level specific, not merely implied.

Trigger Term Quality

"Trigger: diagnose, debug this, something broken/failing/slow" gives several natural phrases a user would actually say. Below anchor 5 because common variants like "not working", "why does X happen", "flaky", "crash", or "regression" are absent; above anchor 3 because the listed terms are natural user language, not jargon.

Distinctiveness Conflict Risk

The niche (a diagnosis discipline for hard bugs and perf regressions) is mostly distinct, with triggers that would not fire for unrelated skills. Below anchor 5 because broad trigger phrases like "something broken/failing/slow" could overlap with testing or code-review skills; above anchor 3 because the framing around hard bugs and perf regressions carves a clear niche.