Skip to content

Research a question

Command: /research-flow · ← All scenarios · User guide

Systematic, grounded investigation tied to your project — the agent first writes the research prompt (which you approve), then runs it and produces a documented answer.

Use this when you need deep investigation or a technology comparison before choosing an approach, grounded in your project’s context.

Not for: simple lookups or single-source questions — just ask the agent directly.

Running it

/research-flow Compare event sourcing vs CRUD for our order service
/research-flow Investigate OAuth 2.0 implementation options for our stack
/research-flow Research vector database options for our RAG pipeline

How it works

The twist is meta-prompting: the agent doesn’t research off your one-line question. It crafts a focused research prompt, shows it to you (because that prompt controls what gets answered), and only then runs a dedicated research pass that weighs multiple options.

flowchart TB
    Start(["/research-flow + question"]) --> Ctx["Load project context"]
    Ctx --> Craft["Craft a research prompt"]
    Craft --> G1{"You approve <br>the research prompt"}
    G1 -->|revise| Craft
    G1 -->|approve| Run["Run the research pass <br>(&ge;3 options, grounded)"]
    Run --> SV{"Self-validation: <br>gaps remaining?"}
    SV -->|yes| Run
    SV -->|no| Final["Finalize the write-up"]
    Final --> Done(["docs/&lt;feature&gt;-research.md"])

    classDef step fill:#dae8ff,stroke:#3674b5,color:#102a43;
    classDef gate fill:#fff2cc,stroke:#b58b00,color:#493800;
    classDef done fill:#e5f7e8,stroke:#35834a,color:#153b20;
    class Ctx,Craft,Run,Final step;
    class G1,SV gate;
    class Start,Done done;

It prioritizes accuracy over speed, compares at least three options, and ends with a self-validation pass. It won’t touch your CONTEXT.md / ARCHITECTURE.md — it only produces the research document.

What you’ll be asked to do

Review and approve the research prompt before the research runs — this is your lever on what the investigation will and won’t cover. Answer any questions it raises about assumptions.

What it creates

plans/<feature>/research-prompt.md (the approved direction) and the final docs/<feature>-research.md, plus a state file.

Analyze a codebase for internal investigation · Write or change code to act on the findings.

Prerequisites: documentation search tools (DeepWiki, Context7).

Sources