The idea

Early AI feels like a manual film camera: powerful, but every focus, exposure, and timing decision is left to the operator.

Agent skills are the semi-automatic modes. A Skill chooses the workflow. A harness holds the process together. Hooks stop predictable mistakes. Verifiers check whether the result earned the claim.

This repository publishes those control layers instead of hiding them. If you are comparing this with a prompt list, start with the one-screen visual proof, then read Control library, not a prompt collection.

The three-minute route

If you do not want to browse 56 folders, follow the product tour: see the control-layer difference, choose one failure to stop, run one Skill in a temporary home, and inspect the matching fixture and verifier.

56 Skills. One control library.

The library now spans engineering quality, research and decisions, agent systems, content and design, and product architecture—without splitting stars, issues, or contributors across small repositories.

Product tour · Start in 60 seconds · Visual proof · Control library, not prompts · Choose a Skill · Core knowledge · Tooling · Browse all 56 Skills · Inspect evidence

Start here

Do not begin by reading 56 folders. Pick the route that matches the failure you want to stop:

If the work fails because... Start with Why
the task is vague or risky dayan-wenzhen creates a falsifiable task contract before polished output hides a wrong assumption
the artifact is hard to explain dayan-deck gives each slide one job and verifies the deck structure
release review is too soft dayan-adversarial-reviewer checks failure modes, maintenance traps, and trust boundaries separately
the repository is unfamiliar dayan-orient maps code before changes start
the agent itself is unclear dayan-agent-designer defines responsibilities, tools, memory, boundaries, and evaluation

The fastest safe trial is a temporary-home install:

git clone https://github.com/Kosmoray/dayan-agent-skills.git
cd dayan-agent-skills
DAYAN_TEST_HOME="$(mktemp -d)"

python3 installers/install.py dayan-wenzhen \
  --agent codex \
  --home "$DAYAN_TEST_HOME"

python3 skills/dayan-wenzhen/scripts/verify_contract.py \
  skills/dayan-wenzhen/examples/starter-contract.json

Generate a paste-ready compatibility report:

python3 scripts/compatibility_smoke.py

Smoke the full library and write a compatibility matrix:

python3 scripts/compatibility_smoke.py \
  --all-skills \
  --json-output docs/compatibility-matrix.json

Run the offline lifecycle smoke:

python3 scripts/runtime_smoke.py

Smoke discovery, trigger routing, example commands, and safe update for the full library:

python3 scripts/runtime_smoke.py \
  --all-skills \
  --json-output docs/runtime-smoke.json

See the full Quickstart, Skill chooser, Compatibility evidence, and FAQ.

Copy-paste examples and fixtures

Try one sanitized run before integrating anything:

The repository now includes 11 sanitized example runs, including:

It also includes 12 artifact fixtures that show copyable output shapes for repository orientation, agent design, guardrail hooks, API review, AI visibility audits, visual deck briefs, architecture decisions, agent packages, responsive UI acceptance, verifier planning, false-positive review, and authority ledgers.

Then use the public playbooks to decide whether your repeated workflow should become a checklist, Skill, verifier, hook, or agent. If you want a bounded contribution, start with Good first issues.

For the system behind the Skills, read the core knowledge map. For the small local tools that verify the repository, read Public tooling.

Choose your route

Create Think Build Verify & Grow
Deck Wenzhen Agent Designer Adversarial Reviewer
Huashu Design Plan Agent Factory AI SEO
HTML Orient Hook Factory
Diagram

All 56 are installable public betas. Featured Skills have dedicated validators; the Core Library uses a shared strict public bundle contract, package-install matrix, and offline lifecycle smoke.

Featured Skills

dayan-deck · public beta

Turn a topic, outline, or source document into a self-contained HTML presentation with:

  • one narrative job per slide;
  • a locked visual and motion system;
  • editable text instead of full-slide screenshots;
  • keyboard, print, responsive, and reduced-motion behavior;
  • a deterministic structural verifier;
  • explicit limits for visual quality, factual accuracy, and PPTX export.

Read the Skill · Open the live starter · Inspect the verifier

dayan-adversarial-reviewer · public beta

Review a concrete change before merge or release through three distinct lenses:

  • failure modes: malformed, repeated, partial, concurrent, interrupted, and rollback paths;
  • maintainability: hidden contracts, mixed responsibilities, and regression traps;
  • trust boundaries: untrusted input, authorization, files, environment, logs, and secrets;
  • evidence-backed BLOCK, CONCERNS, or CLEAN verdicts;
  • matching human-readable Markdown and machine-verifiable JSON;
  • accepted and blocking fixtures plus a deterministic verdict validator.

Read the Skill · Inspect the rubric · Run the validator

dayan-wenzhen · public beta

Turn a vague, risky, or solution-shaped request into a decision-ready task contract before the agent starts generating a polished answer:

  • triage the work type, risk level, currently allowed action, release authority, and minimum evidence;
  • state a falsifiable best-current problem hypothesis rather than pretending certainty;
  • ask only questions whose answers can change the route;
  • compare the current route, alternatives, a reframed third route, and a genuine defer-or-shrink option;
  • end with the smallest reversible bet, pause signal, checkpoint, and evidence needed to continue;
  • return human-readable Markdown plus a machine-verifiable JSON contract.

Read the Skill · Inspect the schema · Run the validator

Install in under a minute

Clone the repository, then install into an explicit agent home. Use a temporary home for first inspection:

git clone https://github.com/Kosmoray/dayan-agent-skills.git
cd dayan-agent-skills
DAYAN_TEST_HOME="$(mktemp -d)"

python3 installers/install.py dayan-deck \
  --agent codex \
  --home "$DAYAN_TEST_HOME"

Replace dayan-deck with any name in the route table to install another public beta Skill.

Claude Code packaging target:

python3 installers/install.py dayan-deck \
  --agent claude-code \
  --home "$DAYAN_TEST_HOME"

The beta installer performs new installs only and refuses to overwrite an existing Skill directory. See the installation contract.

Try the verifier

python3 skills/dayan-deck/scripts/verify_deck.py \
  skills/dayan-deck/examples/starter.html

python3 skills/dayan-adversarial-reviewer/scripts/verify_review.py \
  skills/dayan-adversarial-reviewer/examples/block-review.json

python3 skills/dayan-wenzhen/scripts/verify_contract.py \
  skills/dayan-wenzhen/examples/starter-contract.json

Expected result:

PASS: ...starter.html satisfies the structural deck contract (4 slides)

The verifier deliberately rejects remote runtime dependencies, multiple active slides, missing slide headings, credential-like assignments, private paths, and presenter notes inside audience-facing HTML.

Verify any new bundle

python3 scripts/validate_public_skill.py dayan-orient

python3 scripts/verify_fixtures.py

python3 scripts/verify_methods.py

python3 scripts/verify_positioning.py

python3 scripts/verify_control_demo.py

python3 scripts/verify_share_kit.py

python3 scripts/verify_good_first_issues.py

python3 scripts/verify_tooling.py

See catalog.json for readiness evidence and unresolved host-version evidence for every Skill.

Design principles

  1. Result before ritual. A Skill must create a useful artifact, not merely describe a philosophy.
  2. Evidence before claims. A passing test proves only what the test actually checks.
  3. Open by default. General AI control and verification mechanisms should earn public names, citations, and improvements.
  4. Human authority stays explicit. Publication, spending, credentials, consequential decisions, and high-stakes claims retain human control.
  5. Failure is part of the product. Rejected fixtures and stop conditions matter as much as happy paths.

Compatibility

The repository verifies packaging into Codex and Claude Code Skill directories in isolated temporary homes, then runs an offline lifecycle smoke that discovers installed Skills, routes catalog trigger text, executes declared example commands, and tests marker-guarded update. That is still not a claim that every product version loads the Skill UI or that every model routes identically.

See docs/compatibility.md and docs/runtime-smoke.md.

Contributing

Issues and pull requests are welcome. A useful contribution includes:

  • one repeatable user problem;
  • a clear trigger and non-trigger boundary;
  • one accepted fixture;
  • one rejected or unsafe fixture;
  • a validator or an explicit human-review contract;
  • license provenance.

Start with CONTRIBUTING.md or propose a Skill using the issue template.

Security and provenance

The public beta is a clean-room package. It excludes customer material, private templates, internal infrastructure, machine-specific paths, credentials, and third-party binary assets. Read SANITIZATION.md, SECURITY.md, and each Skill's local provenance and sanitization record.

License

MIT.

Built by Dayan for people who want AI to feel controllable—not magical.