Overview
Zenable automatically analyzes your repositories to maintain high-quality AI context files aligned with your organization’s requirements.How It Works
Scans your repository for AI context files, compares content against your configured requirements, and opens a PR with suggested improvements if issues are found. Files analyzed include:- Claude Code:
**/CLAUDE.md,**/.claude/rules/**/*.md - Cursor:
.cursorrules,.cursor/rules/**/*.mdc(repository root only) - OpenAI Codex:
**/AGENTS.md - GitHub Copilot:
.github/copilot-instructions.md,.github/instructions/**/*.md - Devin Desktop:
**/AGENTS.md,**/.devin/rules/**/*.md - Antigravity CLI:
**/GEMINI.md - Kiro:
**/.kiro/steering/**/*.md
Imported files (@-imports)
Context files often pull in another file with an @path reference—for example, a CLAUDE.md containing @AGENTS.md. Zenable resolves these imports and suggests a shared improvement in the file that actually owns the content (the imported file), instead of duplicating it in every file that references it. When the referenced file is itself a recognized context file, it’s refined in its own right.
A root context file is a recognized context file that is not imported by
another recognized context file. It starts an independent analysis with its
standalone imports included. Each run analyzes at most 12 root context files.
Enabling Analysis
1
Navigate to GitHub Repos
Go to Integrations > GitHub > Repos in zenable.app
2
Select repositories
Click on one or more repositories you want to analyze
3
Open analysis settings
Click the Analysis button to open the configuration modal
4
Enable Refine Context
Toggle on Refine Context to enable AI context file improvements
5
Schedule analysis
Click Add to Priority Queue to run in the next nightly window
When Analysis Runs
Once enabled, analysis runs automatically on a regular schedule.Priority Queue
Click Add to Priority Queue to request to prioritize your repository for a review.Viewing Results
When improvements are found, a PR is created in your repository. Review, modify, or merge as needed.Evaluating Skills
A skill is a folder that holds aSKILL.md file, as the Agent Skills specification defines it. Zenable discovers skills automatically when they sit under a .claude/skills/ or .agents/skills/ folder anywhere in the repository.
Zenable makes sure your skills improve over time by loading the evals/evals.json file from your skill, if one exists. It gives the same tasks to the current skill and to a modified version, runs both on the model your coding agents use most in that repository, and compares the results. If the modified version does the tasks at higher quality, with no regressions and no meaningful increase in token use, Zenable opens a pull request.
You write those tasks in an evals/evals.json file in the skill’s folder. Zenable does not create this file for you.
Context evals start on a paid plan. They also need context refinement turned on for the repository, and Enable deeper context refinement turned on in its configuration profile. Each repository joins after a short settling period.
Add the File
Add oneevals/evals.json for each skill, next to its SKILL.md, and merge it to your default branch. Zenable reads the file from your default branch, so a file on another branch has no effect.
evals/evals.json
File Format
Send this section to your coding agent to make an evals.json for your skill.
File Format
Send this section to your coding agent to make an evals.json for your skill.
The file has these top-level fields:
Each task in
evals can use these fields:A task can be scored only if it has
expected_output, or at least one entry in expectations or assertions. Zenable skips any other task, even if it has output_validations. If no task in the file can be scored, the evaluation ends as Inconclusive and Zenable doesn’t open a pull request for that change.Each entry in output_validations can use these fields:Each text list holds up to 128 entries, and each entry can be up to 8,192 characters. Paths in
files and path must be relative, with no .. and no backslashes.Zenable ignores other keys at the top level and on a task, so fields such as id or name are safe to keep. A key it doesn’t recognize inside an output_validations entry makes the file invalid.If the file isn’t valid JSON, or it breaks a limit on this page, Zenable can’t use it. The change isn’t evaluated and Zenable doesn’t open a pull request for it.Troubleshooting
Analysis hasn't run yet
Analysis hasn't run yet
Verify analysis is enabled and the repository is connected via GitHub integration. Use Add to Priority Queue if you need it sooner.
No PR was created
No PR was created
Your context files are already well-aligned with your requirements—no improvements needed.
Next Steps
- Set up GitHub PR reviews for automated code review
- Define custom policies for your organization