Skip to content
This repository was archived by the owner on Feb 14, 2026. It is now read-only.

kriskimmerle/contextaudit

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 

Repository files navigation

contextaudit

AI agent context quality analyzer. Audit instruction files (AGENTS.md, PROMPT.md, .cursorrules, specs/*) for effectiveness, not just security.

Complementary to agentlint (security) — contextaudit checks whether your instructions will produce good agent output.

Why?

The quality of an AI agent's work depends on the quality of its instructions. Bad instructions → bad results, regardless of model capability. This is especially critical for autonomous loops (Ralph Wiggum pattern, HITL workflows) where the agent runs unattended.

contextaudit checks for the patterns that make agent instructions effective:

  • Completion criteria — Does the agent know when it's done?
  • Backpressure — Are there tests/lints to validate work?
  • Scope boundaries — Is the agent's workspace defined?
  • Task granularity — Are instructions broken into clear steps?
  • Self-correction — Is the agent told to check its own work?

Installation

# Just download it
curl -O https://raw.githubusercontent.com/kriskimmerle/contextaudit/main/contextaudit.py
chmod +x contextaudit.py

# Or clone and run
git clone https://github.com/kriskimmerle/contextaudit.git
cd contextaudit
python3 contextaudit.py

No dependencies. Python 3.9+ required.

Usage

# Audit all instruction files in current project
python3 contextaudit.py

# Audit a specific project
python3 contextaudit.py /path/to/project

# Audit a specific file
python3 contextaudit.py -f AGENTS.md

# Verbose output with suggestions
python3 contextaudit.py --verbose

# JSON output
python3 contextaudit.py --json

# CI mode: exit 1 if grade below B
python3 contextaudit.py --ci --threshold B

# Quiet: just the grade
python3 contextaudit.py -q

Example Output

  contextaudit v1.0.0
  /path/to/project

  Files analyzed: 3
    • AGENTS.md (system)
    • PROMPT.md (prompt)
    • specs/auth.md (spec)

  Grade: A (94/100)

  Issues (3):
    🟡 [C09] [PROMPT.md]: Ambiguous language found: make it better
       💡 Replace with specific, measurable instructions
    🔵 [C07] [AGENTS.md]: No commit/save instructions
       💡 Add: 'Commit after each completed task with a descriptive message'
    🔵 [C05] [specs/auth.md]: No error handling guidance
       💡 Add: 'If tests fail, debug and fix. If stuck, stop and explain.'

  Strengths (7):
    🟢 Has 1 spec file(s)
    🟢 [AGENTS.md] Has backpressure mechanism (tests/lint/build)
    🟢 [AGENTS.md] Well-structured task breakdown (12 steps/items)
    🟢 [AGENTS.md] Has self-correction/verification instructions
    🟢 [AGENTS.md] Specifies concrete tools/commands
    🟢 [PROMPT.md] Has task priority/ordering guidance
    🟢 [PROMPT.md] Follows single-task-per-iteration pattern

  Summary: 1 warning, 2 info

Rules

File-Level Rules (C01-C15)

Rule Severity Description
C01 Critical/Warning Missing completion criteria
C02 Critical No backpressure mechanism (tests/lint/build)
C03 Warning No scope boundaries defined
C04 Warning No structured task breakdown
C05 Info No error handling guidance
C06 Warning Context size issues (too large or too small)
C07 Info No commit/save instructions
C08 Info No self-correction instructions
C09 Warning Ambiguous/vague language detected
C10 Info No specific tools/commands mentioned
C11 Bonus Includes code examples
C12 Bonus Has task priority/ordering
C13 Bonus Single-task-per-iteration pattern
C14 Bonus Subagent/delegation instructions
C15 Info Repeated/duplicate instructions

Project-Level Rules (P01-P04)

Rule Severity Description
P01 Critical No agent instruction files found
P02 Warning No system-level instructions (AGENTS.md)
P03 Bonus Has spec files
P04 Bonus Has implementation plan

Detected Files

contextaudit automatically scans for:

File Type Description
AGENTS.md System Project-wide agent instructions
CLAUDE.md System Claude Code instructions
COPILOT.md System GitHub Copilot instructions
.cursorrules System Cursor editor rules
.windsurfrules System Windsurf editor rules
PROMPT.md Prompt Task-level prompt (Ralph pattern)
PROMPT_plan.md Prompt Planning mode prompt
PROMPT_build.md Prompt Building mode prompt
IMPLEMENTATION_PLAN.md Plan Task plan
specs/*.md Spec Requirement specifications

Grading

Grade Score Meaning
A 90-100 Excellent — agent will have clear, effective instructions
B 80-89 Good — minor improvements possible
C 65-79 Adequate — notable gaps in instruction quality
D 50-64 Below average — agent likely to produce poor results
F 0-49 Poor — critical instructions missing

CI/CD Integration

# GitHub Actions
- name: Check agent context quality
  run: python3 contextaudit.py --ci --threshold B

agentlint vs contextaudit

agentlint contextaudit
Focus Security risks Effectiveness/quality
Checks Injection, exfiltration, destruction Completion criteria, backpressure, scope
Question "Is this safe?" "Will this work well?"
Use together Yes — run both in CI Yes — complementary checks

Testing

# Audit this project's own instructions
python3 contextaudit.py .

# Audit any project with AGENTS.md
python3 contextaudit.py ~/my-project

License

MIT

About

AI Agent Context Quality Analyzer. Audit AGENTS.md, PROMPT.md, .cursorrules for effectiveness. 19 rules, A-F grading, CI mode. Complementary to agentlint.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages