HumanLayer Skills: instruction preservation and reviewable control loops
HumanLayer Skills explained: repository guidance and reviewable agent loops
Understand the five skills, their different modification scopes and why the collection is not a ready-made autonomous maintenance service.
What you will learn
- A collection of instructions with different deliverables
- Select by the change you are willing to authorize
- Treat upstream claims as hypotheses until measured
Before you start
- Repository instructions and basic GitHub Actions concepts
- Understanding of review scope and persistent agent context
Design a bounded, inspectable workflow and distinguish template assumptions from tested behavior.
Key takeaways
- The five skills have different outputs and authority requirements.
- Reference templates are not deployed maintenance services.
- Instruction-adherence claims need their own evaluation.
A collection of instructions with different deliverables
At revision 3c26291, humanlayer/skills lists five Claude Code skills: improve-claude-md, narrow-react-prop-types, build-iterated-agentic-loop, design-control-loop and show-me. These address instruction organization, type contracts, recurring agent work and explanation artifacts. They are not five interchangeable commands for the same task.
The repository provides instructions and reference material rather than a deployed maintenance service. Some skills rewrite a file; others help design scripts, memory and CI workflows for the target repository. Installing a skill therefore does not mean its proposed control loop already exists or has been tested in your project.
Select by the change you are willing to authorize
Improving CLAUDE.md changes guidance future agents may read. Narrowing props can alter an application or public component contract. Designing a loop can eventually add recurring runs and write-capable credentials. Choose a skill by the permitted deliverable, not simply because all three activities sound like code quality improvements.
For example, a request to review repository instructions should not automatically create a scheduled maintenance workflow. Likewise, an explanation artifact from show-me does not authorize changing the code it illustrates. Keep analysis, proposed changes, implementation and recurring operation as separate decisions where their consequences differ.
Treat upstream claims as hypotheses until measured
The instruction-rewrite skill proposes conditional important blocks as relevance signals. Its claims about improved adherence are not a measured result from our review. The control-loop skill proposes a design process and templates; those templates contain placeholders and cannot be treated as a complete production implementation.
This series inspects the pinned README, selected skills, workflow skeleton and iteration helper, with official GitHub documentation for manual triggering. It includes a separate synthetic admission-policy model. No live agent, repository instruction rewrite, scheduled workflow, PR creation or credentialed iteration was executed.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Name the intended deliverable before selecting a skill.
- 2
Separate read-only review from changes and recurring operation.
- 3
Define evidence required before adopting the result.
Copy-ready example
{
"example": true,
"selectedSkill": "improve-claude-md",
"requestedOutput": "proposed rewrite for review",
"applyToRepository": false,
"scheduleWorkflow": false,
"modelTrialExecuted": false
}Frequently asked questions
Does installing the collection create a working maintenance loop?
No. The loop must be designed, implemented and validated for the target repository.
Was instruction adherence improved in a live model test here?
No. This review did not run a model-based adherence experiment.
Sources
- HumanLayer Skills / README.mdSource checked 2026-09-14
- HumanLayer Skills / plugins/improve-claude-md/skills/improve-claude-md/SKILL.mdSource checked 2026-09-14
- HumanLayer Skills / plugins/design-control-loop/skills/design-control-loop/SKILL.mdSource checked 2026-09-14