Atomic Skills
How to give individual skills the right amount of weight in a run
Atomic skills are the small operating contracts inside the harness.
They do one bounded job. A phase wrapper coordinates a lifecycle. An atomic skill changes one move inside that lifecycle: research, handoff, review, writing fragments, shaping a page, checking type quality, or gathering runtime evidence.
The Skill Reference lists what exists. This page explains how to give individual skills the right amount of weight when running the harness.
How Atomic Skills Enter Work
An atomic skill should enter a run through an explicit reason:
- the user names it
- scoped guidance requires it
- a plan assigns it
- a handoff recommends it
- a phase wrapper delegates to it
That explicit trigger matters. A skill changes how the agent works. The operator should be able to see why it was loaded and what output it was supposed to produce.
Parallel Research
Use parallel-research when the next decision benefits from several independent readonly investigations.
Good fits:
- comparing separate code paths before planning
- auditing unrelated features for regression risk
- investigating multiple hypotheses for a bug without editing code
- gathering docs or source evidence before writing a plan
- splitting a broad analysis across independent artifacts
Avoid it when:
- the work requires code edits
- the next step depends tightly on one local investigation
- the question is small enough to answer directly
- delegation would duplicate the main thread's immediate work
The useful output is a synthesis, not a pile of summaries. A good parallel-research closeout says what each worker covered, which evidence is trustworthy, where outputs conflict, and what the next local action should be.
Handoff
Use handoff when a new session should start fresh with proper context before a long, consuming phase.
Good fits:
- requirements are complete and planning should start clean
- a plan is accepted and implementation will be long
- debugging has produced enough evidence for a cleaner resolution run
- implementation is done and docs ingest should begin with curated state
- the context window has accumulated enough stale material to slow the next phase
Avoid it when:
- the agent is midway through the same active phase
- a debugging hypothesis is still being tested
- a coherent edit is half-finished
- a writing pass still has live shape in working memory
The handoff should carry the minimum continuation state: objective, accepted decisions, changed files, verification evidence, open questions, durable artifact links, suggested skills, and anything the next agent should avoid repeating. It should not copy the transcript.
Choosing The Right Weight
Atomic skills deserve different amounts of ceremony.
Some are quick lenses. simplify can be a review pass over a small diff. quality-types can focus a TypeScript boundary check. Others reshape a whole segment of work. parallel-research changes discovery. handoff changes session continuity. The operator should choose the weight by asking how much the skill changes the run, how risky the work is, and what artifact should remain afterward.