Autoresearch Loop

Conducts autonomous, iterative research by defining goals, generating hypotheses, verifying results, modifying approaches, and repeating until criteria are met.

Security Scan

The skill's instructions, metadata, and required resources are internally consistent with an autonomous iterative research assistant and do not request unexpected credentials or installers.

Purpose & Capability

Name/description (autoresearch, iterative verification) matches the SKILL.md loop. The skill requests no binaries, env vars, or installs that would be unrelated to performing research.

Instruction Scope

SKILL.md gives an open-ended methodology (search/analyse/synthesise/verify) and requires verification from multiple sources and iteration caps (max 10). The instructions are high-level and intentionally leave the agent discretion about how to search and which sources to use — this is coherent for a research skill but gives the agent broad authority to query external sources or tools. There are reasonable guardrails (iteration limit, require different approaches, logging), but the skill does not explicitly constrain which data sources or tools may be used or prohibit accessing sensitive files or secrets.

Install Mechanism

Instruction-only skill with no install spec and no code files; nothing is written to disk or downloaded. This is the lowest-risk install profile.

Credentials

No environment variables, credentials, or config paths are requested. The lack of requested secrets is proportionate to a research-oriented skill.

Persistence & Privilege

always is false and the skill is user-invocable; the skill can be invoked autonomously by the agent (platform default) but does not request elevated or persistent privileges. Combined with no requested credentials, the privilege footprint is minimal.

Assessment

This skill appears coherent and low-risk because it is instruction-only and requests no credentials. Before installing or enabling autonomous invocation, consider: 1) Review what tools your agent has (web browsing, external API access, filesystem access) — the skill's instructions allow the agent to use whatever search/synthesis tools it already has. 2) If you don't want the agent to access the web or local files, disable those capabilities or require user confirmation. 3) Require the skill to produce explicit citations for claims and a confidence level (the SKILL.md asks for this — enforce it). 4) Keep the provided iteration limit (10) and consider lowering it if you want tighter control. 5) Don't provide secrets or credentials to the agent while running open-ended research. If you want more assurance, ask the publisher for example sessions or an explicit list of allowed sources/tools; if the skill shipped code or an installer, re-evaluate (that would raise new risks).

Autoresearch Skill

Trigger

Autonomous goal-directed iteration — agent modifies, verifies, keeps/discards, and repeats.

Trigger phrases: "research this thoroughly", "autonomous research", "iterate until complete", "deep dive", "autoresearch"

Core Loop

Inspired by Karpathy's autoresearch methodology:

1. Define goal and success criteria
2. Generate hypothesis or approach
3. Execute (search, analyse, synthesise)
4. Verify result against criteria
5. If criteria met → keep result, move to next
6. If criteria not met → modify approach, retry
7. Repeat until all criteria satisfied

Implementation

# Autoresearch: [Topic]

## Goal
[What you're trying to find/prove/analyse]

## Success Criteria
- [ ] [Criterion 1 — specific and measurable]
- [ ] [Criterion 2]
- [ ] [Criterion 3]

## Iteration Log
### Attempt 1
- Approach: [what was tried]
- Result: [what was found]
- Assessment: [met criteria? why/why not?]
- Next: [what to try differently]

### Attempt 2
...

## Final Output
[Synthesised result that meets all criteria]

Rules

• Always define success criteria BEFORE starting research
• Maximum 10 iterations per research question (prevent infinite loops)
• Each iteration must try a DIFFERENT approach (no repeating failed strategies)
• Log every attempt — the failures are as valuable as the successes
• Verify findings from multiple sources before accepting
• Be explicit about confidence level: high/medium/low for each finding