Autoresearch
Autonomous research loop. Takes a topic, runs web searches, fetches sources, synthesizes findings, and files everything into the knowledge graph. Based on Karpathy's autoresearch pattern. Triggers on: "/autoresearch", "research [topic]", "deep dive into [topic]", "investigate [topic]", "find everything about [topic]", "go research".From its SKILL.md
npx -y skills add onur-gokyildiz-bhi/codescope --skill autoresearchAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 21 stars21 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
3.7 KB, as published. Nobody here has run it
autoresearch: Autonomous Research Loop
You are a research agent. Take a topic, run iterative web searches, synthesize findings, and file everything into the codescope knowledge graph. The user gets structured knowledge nodes, not a chat response.
Research Program
Input: topic from user
Constraints:
- Max 3 rounds of search
- Max 5 web fetches per round
- Stop when: all major angles covered OR contradictions resolved OR max rounds hit
- File everything — don't summarize in chat, put it in the graph
Research Loop
Round 1: Broad Survey
- Decompose the topic into 3-5 distinct search angles
- For each angle: WebSearch with 2-3 queries
- For top 2-3 results per angle: WebFetch the content
- Extract from each source:
- Key claims (with confidence)
- Entities (people, orgs, technologies)
- Concepts (patterns, frameworks, approaches)
- Open questions (what's still unclear)
Round 2: Gap Fill
- Review what's missing or contradicted from Round 1
- Run targeted searches for each gap (max 5 queries)
- Fetch and extract from top results
- Update confidence on claims that are now corroborated or contradicted
Round 3: Synthesis (if needed)
- Only if major contradictions or gaps remain
- Final targeted pass, then proceed to filing regardless
Filing into Knowledge Graph
After each round, file immediately (don't wait until all rounds finish):
For each source found:
knowledge_save(
title: "[Source title]",
content: "[Summary + key contributions]",
kind: "source",
source_url: "[URL]",
tags: ["[topic]", "[subtopic]"]
)
For each entity discovered:
knowledge_save(
title: "[Entity name]",
content: "[Who/what, why relevant to topic]",
kind: "entity",
tags: ["[topic]"]
)
For each concept/pattern:
knowledge_save(
title: "[Concept name]",
content: "[Explanation, how it relates to the topic]",
kind: "concept",
tags: ["[topic]"]
)
Cross-link everything:
knowledge_link(from: "Entity A", to: "Concept B", relation: "uses")
knowledge_link(from: "Source X", to: "Claim Y", relation: "supports")
Link to code if relevant:
# Search for related code in the current project
search_functions(query: "[technology name]")
# If found, link:
knowledge_link(from: "OAuth2", to: "handle_auth", relation: "implemented_by")
Output
After all rounds complete, provide a synthesis report:
## Research Complete: [topic]
**Rounds:** 2 (stopped: all angles covered)
**Sources processed:** 8
**Knowledge nodes created:** 15
- 3 entities
- 5 concepts
- 4 sources
- 3 claims
**Code cross-references:** 4
**Contradictions found:** 1 (flagged)
**Open questions:** 2 (filed for follow-up)
### Key Findings
1. [Most important finding]
2. [Second finding]
3. [Third finding]
### Contradictions
- [Claim A] vs [Claim B] — needs human resolution
All findings filed to knowledge graph. Query with:
knowledge_search(query: "[topic]")
References
See references/program.md for customizable research constraints.