Context diet

Skill jorjani/product-squad/skills/context-diet

Audit always-loaded context — measure consumption of every file loaded each session, classify sections HOT/WARM/COLD, and propose moves out of the hot path. Use when CLAUDE.md / memory files feel bloated and you want to reclaim context budget. Audit and propose only; makes no changes. Distinct from /harness-audit (config health) and /usage-audit (live token burn).From its SKILL.md

Install

npx -y skills add jorjani/product-squad --skill context-diet

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.

SKILL.md

3.2 KB, as published. Nobody here has run it

/context-diet — Context Window Budget Audit

Audit always-loaded context: measure consumption, classify by layer, propose moves.

Usage: /context-diet

For config health use /harness-audit; for diagnosing live token burn use /usage-audit.

Step 1 — Measure

Read and count lines for every always-loaded file:

Global (always): your global CLAUDE.md, your memory index (first 200 lines), any project-root CLAUDE.md, and any always-loaded index/manifest.

Project-specific (when a project is identified): your design-system / project context files (project.json, context.md, schema, config YAML) and the project's CLAUDE.md / .claude/CLAUDE.md.

Active this session: any skill/command definitions loaded.

Estimate tokens: words × 1.3 (English + code heuristic).

Step 2 — Classify

Read every section in each CLAUDE.md and memory file. Apply the three-layer model:

HOT — every interaction, any project. Belongs in CLAUDE.md. Leave it.
WARM — specific projects or workflows. Move to a path-scoped rule or command file.
COLD — reference, examples, detailed instructions. Move to a project doc or delete.

Step 3 — Report

## Context Diet — {date}

### Budget
| File | Lines | Est. Tokens | Status |
| Global CLAUDE.md | X | X | OK / OVER (target <100 lines) |
| Project CLAUDE.md | X | X | OK / OVER (target <150 lines) |
| Memory index | X | X | OK / OVER (target <200 lines) |
| design-system / context files | X | X | {N files} |
| **Total** | **X** | **X** | **~{X}% of window** |

### Findings
{WARM and COLD only; skip HOT. Format: "Lines X-Y: {section} ({N} lines) — WARM/COLD — {where it belongs}". Sort by line count descending.}

### Proposed Moves
- Source: {file}:{lines}
- Destination: {command file, path-scoped rules file, or project doc}
- Savings: {N} lines removed from hot path

### Summary
Current: {X} always-loaded lines (~{X} tokens). After moves: {Y} lines ({Z}% reduction).

Rules

Targets: global CLAUDE.md <100 lines, project CLAUDE.md <150, memory index <200.
Audit and propose only — do NOT make changes.
Skip sections under 5 lines.
Skills are lazy-loaded — list separately; don't count against budget.
Everything in budget and well-placed? "All clear. No action needed."

Deletion Safety

When trimming a file (especially via subagent), diff the result and categorize every deletion:

Safe — exact duplicates, dead file refs, content verbatim in another loaded file, generic advice already in global CLAUDE.md.
Needs review — project-specific behavioral rules that look generic but aren't (e.g. "show ASCII mockup before building"), quickstart recipes, design philosophy gates.

Present both categories to the user before finalizing. Bulk rewrites silently drop behavioral rules that look like filler.

$ARGUMENTS