Content Quality Audit (CQ items)
Two content-level checks complementing the 19 structural items (F1, F2, L1, H1-H3, S1-S11, A1-A2). Source: Anthropic Engineering, "Lessons from Building Claude Code: How We Use Skills" (2026).
Structural items verify that the SKILL.md looks right. Content items verify that the SKILL.md says something the model would not already do.
CQ1 — Obviousness Density
Definition: Proportion of lines that restate default programming knowledge the model already exhibits without the skill.
Why it matters: Obvious content wastes context, dilutes high-signal directives, and trains the model to skim the skill rather than apply it.
Detection signals
A line is flagged as obvious when it matches one or more:
| Signal | Example |
|---|---|
| Generic language tip | "Use const instead of var in JavaScript." |
| Universal good-practice | "Write tests for your code." |
| Tautological framing | "Bugs should be fixed." |
| Restated tool description | "The grep command searches files." |
| Empty hedge | "Be careful when modifying production code." |
A line is not obvious when it:
- Names a project-specific path, identifier, or convention.
- Records a non-default behavior the model would otherwise miss.
- Captures a system-specific quirk (auth flow, env divergence, append-only table).
- Contains a measurable threshold, deadline, or limit.
Scoring
| Density (obvious lines / total non-frontmatter lines) | Verdict | Priority |
|---|---|---|
| < 5% | PASS | — |
| 5% - 15% | PARTIAL | P2 |
| > 15% | FAIL | P1 |
Apply 2-of-3 corroboration (Gauge Core Contract): single-pattern hits enter the soft-flag queue.
Fix template
For each flagged line, propose one of:
- Delete — line carries no project-specific signal.
- Specialize — rewrite with concrete file path, threshold, or system name.
- Promote to Gotchas — if the line records a real trap, move it to the
## Gotchassection.
CQ2 — Description Trigger-Word Presence
Definition: The frontmatter description field is written for model discovery (when to fire this skill), not as a human summary.
Why it matters: The model reads description first to decide whether to load the skill. Human-summary prose ("This skill helps you …") does not anchor activation intent.
Detection signals
A description is trigger-anchored when it contains:
| Signal | Example |
|---|---|
| Activation verb at start | "Use when …", "Trigger when …", "Run after …" |
| Concrete invocation keywords | "babysit", "audit", "rotate", "ship", "rollback" |
| Negative routing ("don't use for") | "Don't use for X (Agent), Y (Agent)." |
| Domain noun list | "PRs, CI status, release notes, changelog." |
A description is under-anchored when it:
- Starts with "This skill …" or "Helps with …".
- Describes the agent's persona without a trigger.
- Lists capabilities in vendor-marketing voice ("powerful", "comprehensive").
- Lacks any of the four signals above.
Scoring
| Trigger-signal count | Verdict | Priority |
|---|---|---|
| 3+ signals present | PASS | — |
| 1-2 signals present | PARTIAL | P2 |
| 0 signals present | FAIL | P1 |
Fix template
Rewrite as:
description: "<gerund/verb activation phrase>. Use when <concrete signal 1>, <concrete signal 2>, or <concrete signal 3>. Don't use for <X> (<OtherAgent>)."Reference exemplar: Architect, Sigil, and Chain descriptions in the current ecosystem.
Audit invocation
gauge audit runs CQ1 and CQ2 alongside the 19 structural items.
gauge checklist CQ1 or gauge checklist CQ2 narrows to a single content item.
Both items respect the change-budget and self-evolution Safety Levels defined in self-evolution.md.
False-positive guards
- CQ1: Skip lines inside fenced code blocks, frontmatter, and
<!-- HTML comments -->. Skip the closing motto line ("> ..."). - CQ1: Do not flag rationale lines that explicitly justify a non-default choice ("Why: …", "Rationale: …").
- CQ2: A description that intentionally omits the trigger because the skill is invoked via Recipe subcommand only (e.g., always routed by Nexus) may waive CQ2 with a top-line HTML comment
<!-- CQ2-waived: Recipe-only invocation -->.
Provenance
These items derive from Anthropic Engineering, "Lessons from Building Claude Code: How We Use Skills" (2026, T1 source). The article identifies "Avoid Obviousness" and "description as model-discovery trigger" as two of the highest-impact principles observed across hundreds of internal skills.