SuperAPI AI Directory

research-codebase

Writing & research

View repository

Updated 10-16 01:35

About this skill

Research Codebase

Triggered by: /research-codebase [questions-file-or-inline-questions] Example: /research-codebase 2026-04-14-SP-42-research-questions.md

Conduct objective codebase research driven by a set of questions. Produce a self-contained research document usable in a fresh context window.

This skill must NOT see the ticket or task description. It receives only research questions. This prevents implementation bias from coloring findings.

Inputs

Research questions - either a file path (output from /generate-questions) or inline questions from the user. If no questions are provided, ask the user for them.

If the user tries to provide a ticket or task description instead of questions, redirect them to /generate-questions first.

Process

Step 1: Read questions and any referenced files

Read the questions file fully. If specific files are mentioned in the questions, read those files in the main context before spawning sub-agents.

Step 2: Decompose into vertical research tasks

Break questions into parallel research tasks. Bias toward vertical slices - trace flows end-to-end rather than cataloging horizontal layers.

Spawn parallel sub-agents:

  • Explorer agents to find where components live
  • Tracer agents to follow a flow from entry point to data store
  • Pattern agents to find how the codebase handles similar concerns elsewhere

Each sub-agent prompt must:

  • Include only its specific questions
  • State: "Document what exists. No opinions, no suggestions, no improvements."
  • Request file:line references for all findings
  • Request actual code snippets for patterns found

Step 3: Synthesize findings

Wait for ALL sub-agents to complete. Then compile results into a self-contained research document.

Step 4: Write output

Write to the working directory: YYYY-MM-DD-TOPIC-research.md

# Research: [Topic Area]

> Self-contained research document. Usable without prior conversation context.

## Questions Investigated
[List of questions this research addressed]

## Summary
[3-5 bullet points answering the core questions]

## Detailed Findings

### [Finding Area 1]
[What exists, how it works, where it lives]
- `path/to/file.ext:L42` - [what this code does]

### [Finding Area 2]
...

## Patterns Found

Reusable patterns discovered in the codebase, with code snippets:

### [Pattern Name]
**Where**: `path/to/file.ext:L42-L60`
**Used by**: [list of consumers]
```[language]
// actual code snippet

[Pattern Name]

...

Data Flow

[How data moves through the relevant systems, traced end-to-end]

Open Questions

[Things that could not be determined from code alone]


## Rules

- **Document what IS, not what SHOULD BE.** You are a cartographer, not an architect.
- **No implementation opinions.** If you catch yourself writing "should", "could be improved", or "consider", delete it.
- **Patterns must include real code.** Every pattern needs a snippet with file:line. Never invent examples.
- **Self-contained output.** The research doc must be usable in a fresh context window with zero prior conversation.
- **Vertical over horizontal.** Trace flows end-to-end. "How does a request become a response" beats "list all middleware."
- **Stay under 40 instructions total.**