We Analyzed 100 AI Coding Configs: Most Are Lean — 10 Files Held 97% of the Waste
TokenCheat Team
7/2/2026
Everyone in AI-coding land repeats the same claim: your instruction files are bloated and burning tokens. We wanted numbers instead of vibes, so we collected 100 real configuration files and ran every one through TokenCheat's deterministic optimizer.
The result surprised us: the median config is lean. The waste story is real — but it's a tail story, not an everyone story.
The corpus
- 100 files: 34
CLAUDE.md, 47AGENTS.md, 19.cursorrules - 93 public GitHub files (root-level only, forks excluded, deduplicated by content hash, most-starred first — top repo: 4,366 stars) plus 7 anonymized local files from our own projects
- Every number below is computed from the raw per-file results
(
results.jsonl); token counts use the same chars÷4 estimate as our free tools
What a typical config looks like
| Metric | Value |
|---|---|
| Median size | 964 tokens (92 lines) |
Median CLAUDE.md | 1,285 tokens |
Median AGENTS.md | 732 tokens |
Median .cursorrules | 1,328 tokens |
| Largest file | 24,186 tokens |
| Over 200 lines | 24% |
| Over 500 lines | 10% |
A ~1,000-token instruction file that's read at the start of every session is a perfectly reasonable spend. If that's you: you're fine. Stop worrying.
Where the waste actually lives
Our optimizer found at least one issue in 34 of 100 files — but the distribution is brutally skewed:
- Total detectable waste across the corpus: 14,530 tokens
- The top 10 files hold 97% of it.
- Files over 200 lines (24% of the corpus) hold 78% of all waste
- Median waste in files ≤200 lines: zero
The worst offenders:
| File | Size (tokens) | Detected waste | Share of file |
|---|---|---|---|
| #1 (64★ repo, CLAUDE.md) | 9,772 | 3,325 | 34% |
| #2 (.cursorrules) | 9,523 | 2,166 | 23% |
| #3 (CLAUDE.md) | 3,139 | 2,073 | 66% |
| #4 (90★ repo, .cursorrules) | 24,186 | 1,690 | 7% |
| #5 (AGENTS.md) | 5,032 | 1,113 | 22% |
The #1 anti-pattern: session logs in your instruction file
The single biggest source of waste we found wasn't duplication or verbosity — it was oversized sections that don't belong in always-on context. 10% of files had at least one section over 600 tokens.
The worst example: a CLAUDE.md carrying dated development-session notes as sections — one of them alone was ~2,965 tokens of "2026-04-24 Follow-up" work-in-progress narration. The agent reads that history on every session, forever.
Instruction files are for durable rules. Logs, decisions, and session notes belong in separate files the agent reads on demand.
The finding breakdown across the corpus:
- Oversized sections (flagged, >600 tokens): 10% of files — 13,348 tokens of potential savings
- Duplicate content (auto-fixable): 16% of files — 1,182 tokens
- File references to verify for staleness: 8% of files
What the tail costs
Straightforward arithmetic on the #1 offender (9,772 tokens, read every
session): at Claude Sonnet 4.6 input pricing ($3/M), 10 sessions/day ×
22 workdays = 2.1M tokens/month ≈ $6.45/month per developer — for one
file. At frontier pricing (Claude Fable 5, $10/M input), **$21.50/month per
developer**. A ten-dev team on a frontier tier pays ~$215/month to re-read
one bloated file.
And the dollars understate it: those tokens crowd the context window on every single turn, which degrades what the agent can actually hold about your code.
Honest limitations
- These are deterministic, static findings only: duplicates, whitespace, filler phrasing, oversized sections, unverifiable file references. We did not judge prose quality, and we could not verify stale file references without each repo's file tree — so these numbers undercount real waste.
- Token counts are chars÷4 estimates, not tokenizer-exact.
- Public GitHub configs skew toward developers who care enough to publish them. Private/enterprise configs may look different — if you run our audit on yours, you'll know.
Check your own config
The same engine that produced these numbers runs free, entirely in your browser — nothing you paste is uploaded:
Run the free stack audit → — paste your CLAUDE.md, get your Context Efficiency Score, the fixes, and a downloadable optimized file.
Or from the terminal: tokencheat discover finds and measures every config
on your machine (CLI download).
Rather have us do it? The done-for-you stack audit covers your full setup — configs, MCP servers, model routing, caching — with the fixes delivered, not just described.