dayan-agent-skills

Control-layer design

Use this when you have a repeated AI workflow and need to decide what to build.

Decision ladder

Problem shape	Build	Why
The same check catches the same failure every time	deterministic script or hook	cheaper, testable, fewer model degrees of freedom
The workflow has a repeatable trigger and artifact	Skill	the model needs judgment, but the boundary is stable
The work spans multiple tools, memory, and handoffs	agent contract	coordination matters more than a single prompt
The output is consequential or public	verifier plus human release	fluent generation must not approve itself

Minimum contract

Write this before implementation:

Repeated problem:
User trigger:
Non-trigger:
Allowed actions:
Human authority:
Artifact:
Accepted example:
Rejected or boundary example:
Verifier or review rule:
Known evidence gap:

Stop conditions

Do not promote a workflow to a public Skill when:

• it only describes one private project;
• the trigger needs private context to make sense;
• there is no accepted example;
• there is no rejected example or boundary;
• publication would expose customer material, credentials, private paths, or internal strategy.

Upgrade path

1. Checklist: stable human process.
2. Skill: stable trigger plus model judgment.
3. Hook or verifier: stable failure pattern.
4. Agent: multiple responsibilities, tools, and state.
5. Marketplace package: install, provenance, sanitization, validation, and contribution route.

This site is open source. Improve this page.