Skip to main content
AutoSearch’s behavior is defined by skills — markdown files in skills/ that the agent reads and follows at runtime. Skills are the “K” in the AVO equation Vary(P_t) = Agent(P_t, K, f).

How skills work

Each skill is a .md file with:
  • YAML frontmatter: name and description (first sentence = when to use)
  • Free-form body: instructions the agent follows
Skills can be modified by the evolution loop — if a change improves the judge score, it gets committed. If it regresses, it gets reverted.

Pipeline skills

SkillPurpose
pipeline-flowOrchestrates the 7-phase pipeline
research-modeChoose speed/balanced/deep from user context
decompose-taskBreak complex questions into sub-questions
gene-queryGenerate diverse queries from entity/pain/object genes
select-channelsPick 5-10 most relevant channels per topic

Knowledge skills

SkillPurpose
systematic-recallExtract everything Claude knows before searching
use-own-knowledgeContribute training knowledge as a source
knowledge-mapSave/load structured knowledge maps across sessions
extract-knowledgePull reusable knowledge from high-quality results

Evaluation skills

SkillPurpose
llm-evaluateSemantic relevance judgment on search results
rerank-evidenceRank results by importance for synthesis
normalize-resultsStandardize results into canonical schema
extract-datesExtract publication dates from all available signals
generate-rubricsDefine binary rubrics for quality evaluation
check-rubricsEvaluate each rubric as pass/fail with evidence

Synthesis skills

SkillPurpose
assemble-contextSelect, deduplicate, compress evidence for synthesis
synthesize-knowledgeTurn evidence into conceptual frameworks
evaluate-deliverySelf-check delivery quality before presenting

Evolution skills

SkillPurpose
auto-evolveRun one AVO evolution step
anti-cheatReject novelty collapse and score gaming
outcome-trackerLearn which queries produced used results
provider-healthTrack platform health and skip cooling-down sources

Meta-skills (immutable)

These five skills cannot be modified by evolution:
SkillPurpose
create-skillCreate new skills when a capability is missing
observe-userLearn user preferences and project context
extract-knowledgePull reusable knowledge from results
interact-userHandle user dialogue and feedback
discover-environmentScan available tools, models, and API keys

Utility skills

SkillPurpose
fetch-webpageFull content extraction from URLs
follow-linksFollow outlinks from curated collections
search-author-trackDiscover a researcher’s full body of work
search-citation-graphFollow citation chains from known papers