Context budget

Skill Geun-Oh/personal-harness/skills/context-budget

Context Budget Monitor — Analyzes context window budget ratios (CLAUDE.md ≤5%, rules ≤3%, MEMORY.md ≤2%, actual work ≥80%) and warns about obesity symptoms. Triggers on 'context budget', 'token ratio'.From its SKILL.md

Install

npx -y skills add Geun-Oh/personal-harness --skill context-budget

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

2.0 KB, as published. Nobody here has run it

Context Budget Monitor

Analyzes context window budget allocation to provide early warning of Context Rot.

Budget Reference Table

Source	Recommended Ratio	Description
CLAUDE.md	≤5%	Table-of-contents role, within 100 lines
.claude/rules/	≤3%	Active rules only
MEMORY.md	≤2%	Within 200 lines
Tool definitions	≤10%	Minimize number of tools
Actual work context	≥80%	Code, tests, error logs

Measurement Method

Estimates approximate token counts for each file (1 token ≈ 4 characters for English, ≈ 1.5 characters for Korean):

Read CLAUDE.md → line count, character count, estimated token count
Read all .claude/rules/*.md → sum totals
Read MEMORY.md → line count, character count
Calculate overall ratios

Obesity Symptom Detection

Warns when any of the following are detected:

CLAUDE.md exceeds 100 lines
MEMORY.md exceeds 200 lines
Total rules files estimated to exceed 3% of context
A single rules file exceeds 50 lines

Output Format

## Context Budget Analysis

### Current Allocation
| Source | Lines | Est. Tokens | Ratio | Status |
|------|------|----------|------|------|
| CLAUDE.md | N | ~T | X% | OK/WARN |
| rules/ | N | ~T | X% | OK/WARN |
| MEMORY.md | N | ~T | X% | OK/WARN |

### Obesity Symptoms
- [WARN] CLAUDE.md 150 lines — reduce to 100 lines or fewer
- [OK] No repeated file-read patterns detected

### Recommended Actions
1. Move detailed content from CLAUDE.md to L1 documents
2. ...

Notes

Token counts are estimates (exact tokenization varies by model)
This check is based on static files. Runtime context consumption is not measured
Integrating with the Hook version (Gate Runner #6) enables real-time monitoring