For skill authors¶

Task-oriented how-tos for building skills, agents, and the runtime around them. The pages are grouped by what you're trying to do — pick the cluster that matches your task.

If you haven't already, work through Getting started first — these how-tos assume you have Reyn installed and have at least skimmed Concepts: phase vs skill vs OS.

Foundation¶

Start here if you've finished the tutorials and want to author a skill yourself.

Write your first custom skill — build a skill from scratch by hand. Walk through skill.md, phases/<name>.md, and artifacts/<name>.yaml with a complete worked example.
Import an existing skill — bring a prompt or another framework's spec into Reyn's DSL using skill_importer.

Composition & multi-agent¶

How skills compose with each other and with peer agents.

Compose skills with run_skill — call one skill from inside another.
Iterate with fan-out — apply a sub-step over a list and collect results.
Build an agent team — set up multiple agents with role-specific skill allowlists.
Multi-hop delegation — chain delegations through more than one agent.
Use plan mode for multi-step tasks — decompose a complex chat request into async steps with crash recovery and operator escape hatches.
Restrict agent skills — scope what each agent in a team can run.

Phase mechanics¶

Deterministic controls inside a phase — pre-LLM steps, schema validation, state persistence.

Add a Python preprocessor — safe vs unsafe mode, signatures, sandbox boundaries.
Validate artifacts — strict-mode checks and schema patterns.
Persist state — what survives across runs and how the workspace stores it.

Reliability¶

Running skills in production — crash recovery, auditing, cost control.

Crash recovery and resume — how the WAL + forward-replay works, how to control resume behaviour, and when to reset. Start here before deploying to production.
Audit and explainability — reading the events log, what it proves, and how to use it for compliance or internal review.

Operations¶

Debugging, integration with external services.

Debug with events — read the JSONL log to understand what happened.
Use an MCP server — wire an external tool into a phase via the mcp Control IR op.

UX & polish¶

User-facing surface area for finished skills.

Author a design — Claude Design integration for visual artifacts.
Localize output — output_language and per-phase locale handling.
Enable voice input — voice-driven chat mode.

Working with stdlib authoring tools¶

Rubrics and reference material for the LLM-driven authoring stdlib (skill_builder, skill_improver, skill_importer, eval_builder). Their primary reader is a Reyn skill, but humans read them too when something goes wrong.

skill-builder checklist
eval-builder rubric
skill-importer mapping
skill-improver criteria
Glossary — vocabulary used across the rubrics.