gioe/token-audit

Token Audit

Scans all skill directories and reports token consumption across five categories: size census, companion file analysis, SQL anti-patterns, redundancy detection, and narrative density.

Step 1: Run the Audit

tusk token-audit

Step 2: Analyze Findings

Review the report across the 4 actionable categories below. For each category, read the flagged skills and compose a concrete description of what should change. Skip categories with no actionable findings.

Companion Files

Look for UNCONDITIONAL loads — these inject tokens on every invocation. For each, describe which skill/file needs a conditional guard added.

SQL Anti-Patterns

Focus on WARN-level items (e.g., SELECT * pulling unnecessary columns). INFO items are advisory — only include them if a clear fix is obvious (e.g., replacing SELECT * with named columns).

Redundancy

Look for tusk setup followed by tusk config re-fetches in the same skill. Duplicate tusk commands across subcommand examples are expected and not actionable.

Narrative Density

For skills with prose:code ratio > 3.0, read each flagged SKILL.md and identify cuttable prose:

• "Arguments" boilerplate — "Parse the user's request to determine..." sections the model doesn't need
• "Integration Points" sections — lists of which other skills reference this one
• Explanatory prose that restates the code — e.g., "What It Shows" sections describing CLI output

Compose a single description per skill listing the specific sections to cut or condense.

Step 3: Create Tasks

Take the combined findings from Step 2 and pass them as input to /create-task. Group related fixes into sensible tasks (e.g., one task per category, or one task for all small skills with the same problem pattern). Let /create-task handle decomposition, deduplication, and insertion.