Context optimization audit

Skill silvio-l/skills/skills/context-optimization-audit

My personal Claude Code skills. Engineering process, context hygiene, and AI-agent maintenance — straight from my .claude directory.

Install

npx -y skills add silvio-l/skills --skill context-optimization-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

0 stars0 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 and optimize Claude Code context bloat across skills, agents, commands, MCP/plugin config, and project instructions. Use when asked to reduce token usage, remove overlap, clean up skills, optimize Claude Code setup, or review loaded context.

SKILL.md

5.4 KB, as published. Nobody here has run it

Claude Code Context Optimization Audit

Audit all loaded context sources and propose what to cut, merge, condense, disable, or move. Do not change anything first. Produce a plan, wait for approval, then apply only approved changes.

Goal

Reduce token usage while preserving behavior. Lower duplication, clearer triggers, fewer competing rules, better global/project/task separation.

Core principle

More instructions is not better. Keep always-loaded context short. Move rarely-used detail into narrowly triggered skills, commands, or reference docs.

What to inspect

Search all of these (do not assume paths are complete — use find, ls, rg):

CLAUDE.md, AGENTS.md
.claude/, .claude/skills/, .claude/agents/, .claude/commands/, .claude/settings.json
.mcp.json, package.json (Claude/MCP-related scripts)
Any agents.json, skills.json, opencode.json, or migrated legacy config files
~/.claude/ (user-level config, if accessible)
Project docs repeatedly copied into prompts or agent instructions

Audit process

1. Inventory

For each context source, capture: type · path · size (lines + chars/4 token estimate) · purpose · trigger · global vs project · recently used or obsolete · overlap · risk · initial recommendation.

2. Find duplication and overlap

Check for:

Same workflow in CLAUDE.md, skills, and commands simultaneously
Old JSON agent definitions alongside current Markdown versions
Broad skills with imprecise triggers
Multiple agents with the same role (planner, reviewer, triage)
MCP instructions duplicated in project docs
Security/branch/release rules copied across multiple files

Group by topic: git workflow · security · architecture · testing · UI/UX · agent orchestration · MCP usage · deployment.

3. Classify each item

See REFERENCE.md for full criteria. Labels: Keep · Condense · Merge · Split · Disable · Delete · Move to project · Move to user · Convert to command · Convert to reference · Lazy-load

4. Check usage evidence

Look for: recent git changes, doc references, active file paths matching the skill, current project stack. Mark uncertain items needs confirmation — do not delete based only on absence of evidence.

5. Produce approval plan

Output this before touching any file:

## Context Optimization Plan

### Summary
- Current estimated context size:
- Largest sources:
- Main duplication areas:
- Highest-impact cleanup:
- Estimated token reduction:

### Inventory
| Type | Path | Lines | ~Tokens | Purpose | Overlap | Recommendation | Risk |
|---|---|---:|---:|---|---|---|---|

### Proposed changes
#### Keep
#### Condense
#### Merge
#### Disable / Delete
#### Move
#### Convert to command / reference

### Risks
### Files that would change
### Files that would be removed or archived
### Rollback plan

### Approval required
No files will be changed until you approve this plan.

6. Apply after approval

Check git status first.
Edit only approved files; do not touch unrelated changes.
Prefer surgical edits — preserve important project rules.
When deleting files, prefer archiving unless deletion was explicitly approved.
Recalculate estimates by re-running the Step 1 inventory against the changed config — measured from the post-change files, not estimated from memory.
Report before/after results using the Final Report format below.

If branch policy is known, follow it. If unknown, do not create or delete branches unless asked.

Cleanup rules (condensed)

Condensing: keep trigger + decision logic + safety boundaries. Remove repeated explanations, motivational text, examples that don't change behavior, long lists replaceable by compact criteria.

Merging: choose the clearer file as target, preserve unique behavior from both, remove the weaker duplicate, update all references.

Deleting/Disabling: confirm obsolete/unused, check references first, document why removal is safe.

Special checks

Markdown vs JSON: if both versions of an agent/skill exist, identify which is authoritative and archive the obsolete one.

Global vs project scope: global = stable personal preferences, universal safety rules, cross-project tooling. Project = architecture, domain terms, branch policy, app-specific rules. Skills/commands = repeatable workflows, optional audits, rarely needed detail.

MCP/plugin: remove instructions already covered elsewhere. Disable unused servers if they add noise. Do not remove active servers just to save tokens.

Final report

## Context Optimization Result

### Changed
### Removed or archived
### Before / after estimate
| Area | Before | After | Reduction |
|---|---:|---:|---:|

### Remaining intentional duplication
### Manual follow-ups
### Verification performed
<!-- Re-run the Step 1 inventory against the post-change state and paste the measured before/after counts here. A prose claim like "reviewed the changes" does not satisfy this field. -->

Non-goals

Do not rewrite application code, change business logic, delete files without approval, remove safety/workflow rules, or pretend token estimates are exact.