Conciseness

The protocol substance (contract table, gates, command table, recovery checks) is dense and earns its tokens, but there is genuine tightening available: multi-sentence external quotations (principlesofchaos, SRE book) re-explain concepts Claude already knows such as error budgets, the "Why this protocol is strict" and mid-run-psychology paragraphs argue rather than instruct, the refuse-to-start list restates the four gates already tabulated, and a version-pinned citation ("ISTQB glossary (V4.7.2)") adds time-sensitive detail. This fits the 3 anchor (mostly efficient, some unnecessary explanation that could be tightened); it is above 2 because no section is pure padding, but below 4 because over-explanation appears in several places, not isolated instances.

Actionability

Guidance is copy-paste ready throughout: per-runner inject/abort commands ("kubectl apply -f <experiment>.yaml" / "kubectl delete -f <experiment>.yaml", "litmusctl chaos run -f", "gremlin attack new --command", "toxiproxy-cli toxic add --type <type> --attribute <name=value> <proxy>"), a fully specified drill contract with exact fields, numeric gate pass/fail conditions, and a complete filled-in output record example. This matches the 5 anchor (fully executable commands covering the common cases).

Workflow Clarity

The four stages are explicitly sequenced ("pre-flight gates, then injection, then the live monitor, then recovery validation") with validation as the backbone: every gate is pass/fail with a defined on-failure action, abort steps require recording timestamp/criterion/signal values, recovery validation runs "regardless of how the run ended", and verdicts feed next actions (NOT_RECOVERED becomes an incident; no next drill from an unrecovered state). Feedback loops are present for this risky-operation skill, matching the 5 anchor; the destructive-operation cap of 3 does not apply because validation is pervasive rather than missing.

Progressive Disclosure

The body is the overview/protocol and the three details that would bloat it are split into real, one-level-deep, well-signaled references: golden-signals.md linked from Gate 3 with its purpose stated, worked-example.md introduced by a section describing exactly what it contains, and anti-patterns.md introduced with an enumeration of all nine entries (verified against the actual files, which exist and match). This matches the 5 anchor; a 4 would require buried or unclearly signaled references, which these are not.

Specificity

The description enumerates concrete, distinct actions in third person ("Sets the go/no-go gates (non-production, healthy baseline, live monitoring, tested rollback), blast-radius bound, abort criteria, inject and abort commands and the recovery check") covering every stage of the protocol with no vague filler. This matches the anchor for multiple specific concrete actions with comprehensive coverage; a 4 would require minor gaps in coverage, which are absent.

Completeness

It explicitly answers both questions: "Use when a chaos experiment is already designed and a fault is about to be injected into a running system" is a concrete when-clause, and "Sets the go/no-go gates (...), blast-radius bound, abort criteria, inject and abort commands and the recovery check" is a clear what. This matches the 5 anchor (both what and when with concrete trigger phrasing); a 4 would need the when to be less explicit, which it is not.

Trigger Term Quality

Natural terms a user would say are present ("chaos experiment", "fault ... injected into a running system", "blast-radius", "abort", "rollback", "recovery"), but common synonyms like "chaos drill" (the skill's own dominant term), "game day", or "resilience testing" are missing. Good keyword coverage with a few natural terms missing fits the 4 anchor; it is not 5 because synonym coverage is incomplete, and not 3 because the present terms are exactly what a user would say rather than generic jargon.

Distinctiveness Conflict Risk

The niche is distinct: it claims only the run/injection phase and explicitly excludes design via "already designed", so it will not trigger for experiment-authoring siblings (the body names chaos-experiment-author and failure-injection-test-author). Triggers like "fault about to be injected" are specific to this skill with minimal conflict risk, matching the 5 anchor; the 4 anchor's "minor overlap risk" is already addressed by the already-designed scoping clause.