Atomr Agents Ai Skills

ai-skills/

Skills for AI coding assistants working on projects that depend on atomr-agents — not for editing atomr-agents itself. They follow the standard SKILL.md + frontmatter convention used by Claude Code, Claude Agent SDK, and other agentic tools.

These skills are deliberately separate from the repo's own dev tooling so that distributing them to consumers does not entangle atomr-agents' internal development workflow.

What's here

Skill	Use when…
atomr-agents-quickstart	Standing up the first agent — picking feature flags, wiring Pipeline, building an Agent, running against MockRunner
atomr-agents-pipeline	Composing Callables with Pipeline + decorators (with_retry, with_fallbacks, with_config, with_timeout, Branch)
atomr-agents-state	Channelled state — picking reducers, declaring StateSchema, persisting via Checkpointer, fork-with-edit
atomr-agents-hitl	Human-in-the-loop — interrupt(), static breakpoints, Command::{Continue, Resume, Update, Goto} resume
atomr-agents-rag	Building a retriever pipeline — picking from the zoo, ingesting docs, writing to LongStore
atomr-agents-tools	Authoring tools — Tool vs RichTool, ToolReturn, parallel dispatch semantics, HandoffTool, memory tools
atomr-agents-middleware	Wrapping an agent's per-turn pipeline with AgentMiddleware hooks
atomr-agents-multi-agent	Org / team / department topologies, the four routing patterns, swarm_loop, HandoffTool
atomr-agents-eval	Eval suites, judge / pairwise / rubric scorers, RegressionGate, AnnotationQueue
atomr-agents-observability	EventBus, RunTreeBuilder, tracers (Stdout, Jsonl, LangSmith)
atomr-agents-troubleshooting	Debugging atomr-agents-flavored errors — BudgetExceeded, PolicyDenied, parser failures, retry exhaustion, channel mismatches
atomr-agents-langgraph-migration	Mapping LangChain / LangGraph idioms onto atomr-agents — concept table + concrete code translations

Each SKILL.md is a thin router: it points at canonical docs in this repo (docs/*.md, examples/*) and at the relevant crate's API. It deliberately does not restate API surfaces that belong in rustdoc, because those drift faster than docs.

Installing

Pick the path that matches your assistant. The skills themselves are vendor-neutral SKILL.md files — only the install mechanism differs.

Claude Code (recommended: marketplace)

If you use Claude Code, install via the plugin marketplace — this keeps the skills updated as atomr-agents releases, with no manual copy step:

/plugin marketplace add rustakka/atomr-agents
/plugin install atomr-agents-ai-skills@atomr-agents

You can also install from a local checkout (useful when developing against an atomr-agents fork):

/plugin marketplace add /path/to/atomr-agents
/plugin install atomr-agents-ai-skills@atomr-agents

Skills auto-activate based on the description frontmatter — no need to invoke them explicitly.

Claude Agent SDK / project-local .claude/skills/

For SDK-based agents or project-local Claude Code setups that read from .claude/skills/, copy or symlink the skills in:

# copy (snapshot)
cp -r ai-skills/skills/* .claude/skills/

# symlink (track upstream)
ln -s "$(pwd)/ai-skills/skills/"* .claude/skills/

Stylistic conventions

These skills follow atomr's:

1. SKILL.md + frontmatter — name, description. The description is what triggers auto-activation, so it's specific about when to invoke.
2. Mental model first. Each skill opens with a one-paragraph mental model of the subsystem before diving into API.
3. Working code blocks. Snippets compile against the published crate version; copy-paste is the intended use.
4. Pointer to canonical docs. Every skill ends with a "Canonical references" list — paths to docs/*.md, examples/*, and the relevant crate.
5. "Common mistakes" coda. Failure modes the framework's architecture makes possible (channel-write to wrong key, missing reducer, mid-handle ask, etc.).

Authoring a new skill

ai-skills/skills/atomr-agents-<topic>/
└── SKILL.md

SKILL.md frontmatter must include:

---
name: atomr-agents-<topic>
description: Use when … . Triggers on … .
---

Keep skills focused. If you find yourself documenting two unrelated subsystems, split them. The atomr-agents-troubleshooting skill is the only deliberately multi-topic one — its job is "given an error, where do I look?".