Context window audit

Skill mattdweigand-sketch/agent-skills/skills/context-window-audit

Portable agent skill library for Codex, Claude Code, and other AGENTS-aware tools. Covers eval loops, research, writing, and project hygiene.

Install

npx -y skills add mattdweigand-sketch/agent-skills --skill context-window-audit

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Audit an agent setup for token waste, context-window bloat, stale instructions, and dead command wrappers. Use when the user says context-window audit, context audit, audit my context, check my settings, why is Codex so slow, token optimization, or asks why startup context is too large.

SKILL.md

2.4 KB, as published. Nobody here has run it

Context Window Audit

Use this skill when the user asks for a context-window audit, context audit, token optimization, settings cleanup, or an explanation of why an agent session is loading too much context.

Scope

Audit prompt and context surfaces that load automatically or are likely to be pulled into many sessions:

global and repo instruction files such as AGENTS.md, CLAUDE.md, command wrappers, settings, and skill descriptions
enabled skills, plugins, and connectors when visible
memory summaries or routed context files when they affect startup behavior
repeated references that cause large files to be read by default

Do not audit private content for substance unless it is relevant to context loading. The goal is smaller, clearer context, not content cleanup.

Load references/context-surface-taxonomy.md when classifying issues or formatting the final audit. That file owns the stable action taxonomy and output template.

Process

Identify the startup path.
- List the files, skills, commands, memories, or settings that are loaded or strongly routed at session start.
- Separate automatic context from on-demand references.
Find bloat and dead routes.
- Look for duplicated rules, stale model-specific instructions, wrappers that invoke missing skills or commands, broad "read everything" guidance, and large reference files that should be loaded only on demand.
Classify each issue. Use the action taxonomy in references/context-surface-taxonomy.md.
Report findings before edits.
- Prioritize issues by token impact and failure risk.
- Name the canonical owner for every duplicated rule.
- Call out any command wrapper that points to a missing target.
Apply changes only if asked.
- Keep edits narrow.
- Preserve repo-specific authority boundaries.
- Validate any edited skill frontmatter or command wrapper if a validator is available.

Output

Use the output template in references/context-surface-taxonomy.md.

If no meaningful bloat is found, say that directly and list the remaining context surfaces that still deserve periodic review.