Skill library · 34 packs
Load a skill, not the atlas.
Anthropic-format SKILL.md files distilled from the admitted corpus. Name and description are the index. The body is a playbook. Agents GET the index, then one pack.
GET /api/canon/skills · GET /api/canon/skills/{id}?format=md
Meta
How to write skills, constitutions, and lessons from traces.
- Write a skillExpertise should load sometimes, not always. The constitution (AGENTS.md) is too coarse; a prompt dump is too fat.skill-creator · 10 min
- Write AGENTS.mdA coding agent will touch this repository. The constitution is short, true, and runnable.agents-md · 8 min
- Distill a traceYou have a trace. The lesson is sitting in tokens you already paid for.distill-traces · 8 min
Doctrine
12-factor, workflows vs agents, the loop, framework choice.
- Twelve-factor agentsSomeone will pay, wait, or sue. A framework got you a demo. The last 20% is software engineering.twelve-factor-agents · 12 min
- Workflows versus agentsYou are about to pick a topology. Most production systems that work are workflows with a small agentic island.workflows-vs-agents · 6 min
- The agent loopYou need a loop you can pause, inspect, eval, and resume. Frameworks wrap this; they must not own it.agent-loop · 8 min
- Pick a frameworkThe raw loop works and you still need one primitive. Not before evals. Not because a README went viral.framework-choice · 6 min
Context
The window as a product. Compaction as policy.
ACI & protocols
Tools, permissions, MCP, A2A, humans, sandbox.
- Design the ACIYou are adding or reviewing tools. The ACI is the prompt you cannot see.tool-design · 10 min
- Allow / ask / denyThe agent can touch a workspace, a browser, or production. The model proposes; code decides.permissions · 8 min
- MCP serverA capability must live outside the agent process and be approved independently.mcp-server · 8 min
- Agent-to-agent (A2A)Agents from different runtimes need a contract. This is not a CrewAI crew.a2a-protocol · 6 min
- Humans as toolsA side effect needs a person, or the agent is missing information only a person has.human-in-the-loop · 8 min
- Sandbox and injectionThe agent has a shell, a browser, documents, or plugins. The model is a confused deputy.sandbox-security · 8 min
Patterns
Chains, routers, crews, harnesses, plan/act.
- Prompt chainYou can name the stages and check the intermediates.prompt-chaining · 5 min
- Router / handoffStable categories and a cheap, accurate classifier.routing · 5 min
- Fan-out / reduceIndependence is real. Shared mutable state makes this a race, not a pattern.parallelization · 5 min
- Evaluator–optimizerClear scoring criteria and iterative gains. If you cannot write the rubric, you cannot loop it.evaluator-optimizer · 6 min
- Code-actThe agent would otherwise make 40 tiny tool calls to filter a table.code-act · 6 min
- Orchestrator–workersSubtasks are not known up front. The parent must not drown in search hits.orchestrator-workers · 8 min
- Multi-agent topologiesSpecialists need different tools or a fresh window, and a parent must synthesize.multi-agent · 10 min
- Spawn subagentsProtect the parent window, parallelize, or apply a different skill set.subagents · 8 min
- Plan / actSide effects need a human-agreed shape. Plan is a file, not a thought.plan-and-act · 6 min
- Spec-driven workYou can score the work. The agent would otherwise wander.spec-driven · 6 min
- Design a harnessThe product is a desk, not a chat wrapper. Claude Code, Codex, OpenHands, Cline, Goose, Deep Agents converge here.harness-design · 10 min
Jobs
Coding, research, browser, documents, support.
- Coding agentThe product is software engineering — PRs, tests, and a desk, not a chat wrapper.coding-agent · 12 min
- Research agentThe path is unknown, sources are many, and the deliverable is a brief.research-agent · 10 min
- Browser agentThe environment is a web page that has no API you can call.browser-agent · 8 min
- Document / RAG agentUsers ask questions over private documents and you must not hallucinate a missing page.document-agent · 8 min
- Support / ops agentInbound language must become a structured action, with approval on side effects.support-agent · 10 min
Production
Evals, durable execution, traces.
- Eval harnessYou need to know whether a prompt, a tool, or a skill got worse.eval-harness · 10 min
- Durable executionA demo works on your laptop and a customer will wait on it.durable-execution · 10 min
- Trace the loopYou cannot improve what you cannot see. The loop must be inspectable.observability · 6 min