native-agent-skill-validation
Skill5 starsChanged 2 months agoMITView on GitHub
Test whether a Codex skill triggers appropriately or changes behavior using fresh controlled agent trials. Use for behavioral evaluation, not routine skill editing.
In this repository10 other files it gives its agents.
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 8
Use the runtime-provided skill-creator for editing, package structure, and basic validation.
- Distinctive: line 22
Before using agents, load native-subagents-first for lifecycle, model selection, ownership, and budgets.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Native agent skill validation
Use the runtime-provided skill-creator for editing, package structure, and basic validation. This workflow evaluates a candidate that already exists.
Choose the claim
- Natural trigger: expose the candidate metadata with a realistic request that does not name the skill. Distinguish loaded, read, announced, and applied evidence. This tests discovery and usability, not the effect of a revision.
- Revision effect: compare isolated prior and candidate revisions under the same prompt and worker conditions. Verify which skill body each worker actually read and applied.
- Explicit invocation: a named-skill task can show invoked usability. It cannot establish natural discovery or a revision's benefit by itself.
Define pass, fail, and inconclusive outcomes before running. Use a task and observable that distinguish the intended behavior from plausible alternatives. If every alternative passes, the test does not resolve the claim.
Build the task from raw artifacts and withhold the intended answer, suspected defect, proposed fix, and prior conclusions. Keep side effects within the authorized evaluation scope.
Run the relevant study
Before using agents, load native-subagents-first for lifecycle, model selection, ownership, and budgets. Before a controlled study, read controlled trials for isolation, assignment, evidence capture, scoring, and promotion requirements.
Hold worker and environment conditions fixed apart from the named treatment. Use fresh children and uncontaminated workspaces. A mismatched model, prompt, permission, artifact, or policy read makes attribution inconclusive.
Preserve complete answers and verify their cited evidence in the parent. Never treat a completed status, a wording match, or a frontmatter check as proof of useful behavior. Use the typed rollout exporter for outbound evidence; never transmit a source rollout.
Report the exact claim and branches tested, trial conditions and evidence, parent verification, result, confounds, and promotion limits. A single matched pair is exploratory; a behavior-promotion claim requires at least two independent matched trials per condition. Send demonstrated instruction defects back for editing, then test the changed candidate with fresh trials.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

