| name | autoresearch |
| description | Automatically research and reverse-engineer a technique, visual effect, algorithm, or codebase. Given a reference (image, video, URL, library name, description, or existing code), Claude researches how it works, reverse-engineers the approach, and produces a structured findings document. Use when you want to understand "how is this made", "reverse engineer this effect", "research this technique", or "figure out how X works so I can implement it".
|
Autoresearch
Given a reference target, automatically research it and produce a reverse-engineering
report with implementation guidance.
Input (via $ARGUMENTS or conversation context)
The target can be any of:
- A visual reference — image path, screenshot, video frame
- A URL — article, repo, shader toy link, paper
- A library or tool name — "how does pixelmatch work"
- A description — "that liquid metal refraction effect"
- An existing codebase or file — reverse-engineer undocumented code
- A combination of the above
Research phases
1. Identify the target
Clarify what's being researched: a visual technique, an algorithm, an architecture, a
library internals. If ambiguous, ask one focused question before proceeding.
2. Collect evidence
- Visual refs: analyze images/screenshots with vision tools — describe what's happening geometrically, mathematically, temporally
- Code refs: read source files — identify core algorithms, data structures, key functions
- Web research: search for papers, shader implementations, blog posts, prior art
- Existing project context: scan
refs/, docs/, reference/, CLAUDE.md, README.md in the working directory for related material already gathered
3. Reverse engineer
Break the technique into layers:
- What is the high-level effect or behavior?
- What mathematical/algorithmic primitives drive it?
- What are the key parameters and how do they interact?
- What are the non-obvious parts (the "tricks")?
- What constraints apply (platform, performance, compatibility)?
4. Implementation map
Translate findings into actionable implementation notes for the current project:
- Language/platform fit (GLSL ES 1.00, Metal MSL, TypeScript, etc.)
- Suggested implementation order (scaffold → core → details)
- Known pitfalls from the research
Output
Save a Markdown report to refs/research-<topic>-<YYYY-MM-DD>.md if a refs/ folder
exists, otherwise to the current directory as research-<topic>-<YYYY-MM-DD>.md.
# [Topic] — Autoresearch
_Date: YYYY-MM-DD | Source: [reference used]_
## What it is
One paragraph: the effect/technique/system in plain terms.
## How it works
### Core mechanism
### Key mathematical primitives
### Parameters & controls
### The non-obvious parts
## Prior art & references
- [link or file] — what it contributes
## Implementation map
### Platform notes
### Suggested approach
### Pitfalls
## Open questions
- [ ] ...
Behavior
- Identify the target from
$ARGUMENTS or conversation context.
- Run all relevant research phases in parallel where possible.
- Synthesize into the report structure above — omit empty sections.
- Save the file and confirm path to user.
- Optionally: if the project has a
todo.md or tasks.md, offer to append
a follow-up implementation task.
$ARGUMENTS
