Context Bloat
Persistent instructions become too large to stay focused.
Research-defined taxonomy · local screen
Find configuration smells in AGENTS.md.
Screen AGENTS.md or CLAUDE.md for six documented failure patterns. Your file stays in this browser; only anonymous action counts are recorded.
6
smell types
91%
study sample hit
0
file uploads
The 91% figure is from a 100-repository research sample, not a claim about all repositories.
Paste the instruction file
Markdown only. Nothing is sent to a server for analysis.
GitHub: open the file → History. This assesses Init Fossilization.
Public evidence
See the same screen across 39 pinned AGENTS.md files.
Explore repository-level results, exact source links, and JSON or CSV downloads.
The catalog
Six ways instruction files quietly decay.
A 2026 study derived the taxonomy from 14 practitioner sources and evaluated it on 100 popular repositories containing AGENTS.md or CLAUDE.md. The authors found at least one smell in 91 files.
Read the research paperSkill Leakage
Rare procedures occupy context in every session.
Lint Leakage
The file repeats rules already enforced by automation.
Blind Reference
A linked file appears without purpose or scope.
Init Fossilization
Generated instructions are never revisited.
Conflicting Instructions
Rules point the agent in opposite directions.
What this checker can—and cannot—claim
Deterministic and private
The screen runs entirely in your browser and explains each match. Context Bloat and Init Fossilization use the paper’s published thresholds.
An early-warning screen
The paper used language models plus manual review for four smells. This free tool uses disclosed local proxies, so it can produce false positives or miss context-dependent problems.