CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

context-audit

Token-hygiene audit of the always-loaded context stack — CLAUDE.md, AGENTS.md, auto-memory MEMORY.md, and the bootstrap-rendered identity files (SOUL.md, USER.md, ACCESS_POLICY.md, HEARTBEAT.md) or their harness equivalents. Finds redundancy, contradictions, stale content, compression candidates, and skill-extraction candidates; produces a ranked action list sorted by token savings with a risk class per finding. REPORT-ONLY: this skill never edits any audited file. Recommendations for bootstrap-rendered files target the interview answer bank / templates, never the rendered output. Judging routes through `gbrain eval cross-modal` (single cheap model by default; full multi-model panel is explicit opt-in).

Quality

81%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Passed

No findings from the security scan

SKILL.md
Quality
Evals
Security

Quality

Content

81%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

A well-sequenced, highly actionable body: the procedure has an explicit validation gate with exit-code-driven feedback loops, and all commands are copy-paste ready. The main weakness is conciseness — the report-only and rendered-file rules are each restated across four-plus sections, ironically the same redundancy pattern the skill audits — and progressive disclosure could be improved by moving the output template and dedup details into reference files.

Suggestions

State the report-only and rendered-file rules once in the Contract and have other sections link to it ("see Contract") instead of restating them in Procedure, Anti-Patterns, and Dedup — this targets the lowest dimension, conciseness.

Move the full report template (Output Format) and the Dedup section into a reference file (e.g. references/output-format.md) with a one-line pointer, reducing the always-inlined body and improving progressive disclosure.

Trim the motivational 'What this is' paragraph to one sentence; the per-turn cost framing is evident to Claude from the procedure itself.

DimensionReasoningScore

Conciseness

The body is mostly efficient — no concept explanations Claude wouldn't know, and sections like Scope and Output Format earn their tokens — but the same rules are restated multiple times: report-only appears in the Contract ("No audited file is edited, no page is written, nothing is auto-fixed"), Procedure step 5, Anti-Patterns ("Report-only — even 🟢 zero-risk deletions are recommendations"), and Dedup; the rendered-file rule likewise repeats across Contract, Procedure step 4, Anti-Patterns, and Dedup, plus the motivational "per-turn tax" paragraph. This matches the level-3 anchor (mostly efficient but could be tightened): it is above level 2 because there is no padding or explained background material, and below level 4 because the cross-section repetition of two core rules is more than a minor trim.

3 / 5

Actionability

Guidance is fully executable: a copy-paste enumeration loop with the exact token arithmetic ("bytes/2.8 ... (n*10+27)/28"), a complete `gbrain eval cross-modal` invocation with every flag resolved (JUDGE=$(gbrain config get models.tier.utility) with a fallback note via `gbrain models`), enumerated exit codes with per-code responses, and a concrete report template. This matches the level-5 anchor (copy-paste ready commands covering the common cases); level 4 is excluded because there is no missing key detail — even the judge-model resolution and the error paths are specified.

5 / 5

Workflow Clarity

The five-step Procedure is clearly sequenced (enumerate → analyze → classify → judge → deliver) with an explicit validation gate: step 4 defines exit codes "0 PASS — deliver. 1 FAIL — fix the flagged weaknesses in the draft ... and re-judge. 2 INCONCLUSIVE ... deliver the report but label it 'unjudged'", which is a full validate-fix-retry feedback loop, and the 'When it fails' section adds error-recovery guidance for provider/key failures. This matches the level-5 anchor (explicit validation steps, feedback loops for error recovery); level 4 is excluded because no checkpoint is implicit — even the INCONCLUSIVE path has a defined action.

5 / 5

Progressive Disclosure

Good structure with clear sections and well-signaled, one-level-deep references to sibling material ("[conventions/brain-first.md](../conventions/brain-first.md)", "[skills/soul-audit/SKILL.md](../soul-audit/SKILL.md)", the cron-scheduler and brain-ops skills), each with its purpose stated. It sits at level 4 rather than 5 because everything lives inline in a single ~220-line SKILL.md with no bundle split — the Output Format template, the multi-model opt-in detail, and the Dedup section are all candidates for reference files — so a reader consumes the whole body to reach any part; it is above level 3 because the references that do exist are clearly signaled and the body is easy to navigate via headers.

4 / 5

Total

17

/

20

Passed

Description

75%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

A highly specific, third-person description with concrete actions, named audit subjects, and explicit safety boundaries that minimize conflict risk with adjacent skills. Its main weakness is the missing "Use when..." trigger clause inside the description itself, which caps completeness, and slightly thinner natural-language trigger coverage than the frontmatter triggers list provides.

Suggestions

Append an explicit trigger clause to the description, e.g. "Use when the user asks for a 'context audit', 'context diet', 'system prompt audit', 'prompt compression', or says their 'context is too big'" — this lifts completeness from 3 to 5.

Fold 1-2 natural user phrasings ("context is too big", "system prompt audit", "reduce context size") into the description body so trigger-term quality matches the frontmatter triggers list instead of relying on it.

Consider trimming the judging-cost detail ("single cheap model by default; full multi-model panel is explicit opt-in") from the description, as it adds tokens without helping trigger or discovery decisions.

DimensionReasoningScore

Specificity

The description lists multiple concrete, distinct actions — "Finds redundancy, contradictions, stale content, compression candidates, and skill-extraction candidates; produces a ranked action list sorted by token savings with a risk class per finding" — and names the exact files audited (CLAUDE.md, AGENTS.md, MEMORY.md, SOUL.md, USER.md, ACCESS_POLICY.md, HEARTBEAT.md). Coverage is comprehensive and stated in third person ("Finds", "produces"), matching the level-5 anchor; the level-4 anchor is ruled out because no capability area is left as a minor gap.

5 / 5

Completeness

The "what" is clear and detailed (audit the always-loaded stack, find redundancy/contradictions/staleness, produce a ranked savings-sorted action list with risk classes), but the description contains no "Use when..." clause or equivalent explicit trigger guidance, which caps completeness at 3 per the judging guidelines. It is not a 2 because the "what" is fully explicit rather than vague, and not a 4/5 because the "when" is entirely absent from the description text (the frontmatter's separate when_to_use field is outside the evaluated target).

3 / 5

Trigger Term Quality

Good natural keyword coverage: "Token-hygiene audit", "context stack", "compression candidates", "redundancy", "contradictions", "audit" — phrases a user would plausibly say. It falls short of the level-5 anchor because common phrasings users actually say ("context is too big", "system prompt audit", "audit my context stack") appear only in the frontmatter triggers list, not in the description itself; it is above level 3 because several relevant natural terms are present, not just one or two generic ones.

4 / 5

Distinctiveness Conflict Risk

The description carves out a clear niche — "Token-hygiene audit of the always-loaded context stack" — and actively disambiguates from adjacent skills via "REPORT-ONLY: this skill never edits any audited file" and "Recommendations for bootstrap-rendered files target the interview answer bank / templates, never the rendered output". Conflict risk with editing, tuning, or general review skills is minimal, matching the level-5 anchor; the level-4 anchor's "minor overlap risk" is not evidenced since the scope and non-mutating contract are explicitly stated.

5 / 5

Total

17

/

20

Passed

Validation

87%

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 14 / 16 Passed

Validation for skill structure

CriteriaDescriptionResult

frontmatter_unknown_keys

Unknown frontmatter key(s) found; consider removing or moving to metadata

Warning

relative_links

Relative link issues: 4 suspicious

Warning

Total

14

/

16

Passed

Repository

garrytan/gbrain

Path

plugin-variants/gbrain-coding/skills/context-audit/SKILL.md

Reviewed

Table of Contents

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.