dayan-agent-skills

Control library, not a prompt collection

Dayan Agent Skills is not a folder of clever prompts. It is a public control library for making probabilistic AI work more repeatable, inspectable, and safe to hand off.

For the 10-second visual version, open the control-library demo.

The difference

Prompt collection	Control library
Starts from what to say to a model	Starts from the workflow failure to control
Optimizes for impressive output	Optimizes for repeatable decisions and evidence
Trusts the model to remember constraints	Puts repeated constraints into Skills, fixtures, hooks, and verifiers
Treats a good answer as success	Separates artifact ready, internally verified, human accepted, released, externally observed, and economically validated
Hides unsafe boundaries in long instructions	Names authority, stop conditions, and non-claims beside the artifact
Grows by adding more prompts	Grows by adding reusable contracts, examples, tools, tests, and contribution paths

Five layers

1. Skill — when to use the workflow, when not to use it, what artifact to create, and where authority stops.
2. Method — the reusable judgment rule behind a family of Skills.
3. Fixture — a public, copyable example of what useful output looks like.
4. Verifier — a deterministic or checklist-based way to test the parts that should not depend on taste.
5. Evidence boundary — what the repository has proved and what it has not proved yet.

A concrete example

dayan-wenzhen does not only say “ask better questions.” It packages a task-contract workflow:

• input risk and authority are named before solution work starts;
• the current problem hypothesis must be falsifiable;
• alternatives and defer/shrink options remain visible;
• high-risk actions require review boundaries;
• the starter contract is checked by skills/dayan-wenzhen/scripts/verify_contract.py.

That is a control layer. A prompt collection would usually stop at a better instruction.

What to inspect first

• Choose a Skill if you know the failure you want to stop.
• Core knowledge if you want the method layer.
• Artifact fixtures if you want concrete output shapes.
• Public tooling if you want the verifier layer.
• One-screen control demo if you want the visual positioning first.
• Compatibility matrix and runtime smoke if you want package evidence.

What this still does not claim

This repository does not claim universal host-version compatibility, model decision quality, legal clearance, customer adoption, external popularity, or economic outcomes. It claims a public, installable, sanitized, and increasingly verified control surface.

Why one repository

The useful unit is the library: one catalog, one installer, one issue tracker, one release gate, one set of examples, one Pages front door, and one contributor path. Splitting each Skill into a small repository would make the system harder to evaluate and would dilute the public signal.

This site is open source. Improve this page.