Skip to main content

​
What it does

/research runs a structured, multi-source research process that produces sourced findings with explicit confidence levels. Rather than acting on assumptions or training-data knowledge, it gathers current information and validates it before handing findings to a downstream workflow.

​
When to use

Use /research before implementing with unfamiliar technology, validating an approach, comparing options, or answering questions that need current information. When not to use: If the answer is in the existing codebase, read the code first. /research is for unknowns that require external sources.

​
Prerequisites

Planning Mode must be active in Antigravity settings. Deep research requires extended reasoning to evaluate source quality and detect contradictions across multiple sources.

​
Conversation mode

Planning Mode — required.

​
What happens

1

Session announcement

Antigravity outputs a session header listing the specific questions to be answered and what decisions depend on each answer.
2

List questions explicitly

Before any searching, all questions are stated explicitly. Each question is tied to a downstream decision it will inform.
3

Load the deep-research skill

The deep-research skill runs its full 5-step process: source discovery, source quality ranking, claim extraction, cross-source corroboration, and confidence assignment.
4

Apply quality gate

The deep-research quality checklist must score 4/5 before findings can be acted on. If it scores below 4/5, more sources are gathered. Low-confidence findings are never handed off to /brainstorm, /plan, or /design.
5

State decision alignment

Before handing off, Antigravity explicitly states: “Decision: [what I’m about to decide]. Finding that supports it: [key finding + confidence]”.
6

Hand off to downstream workflow

Findings are passed to the appropriate workflow: /brainstorm for feasibility, /plan for architecture confirmation, /design for options evaluation, or directly to you if you asked a question.

​
Skills invoked

  • deep-research — 5-step multi-source research process with source quality ranking

​
Example

/research Is Prisma ORM suitable for a Node.js app with 50k daily active users?
Antigravity outputs: 
Session: /research
Questions: Is Prisma ORM suitable for 50k DAU Node.js app?

Research complete.

Question: Is Prisma ORM suitable for a Node.js app with 50k daily active users?
Finding: Yes, with connection pooling (PgBouncer) required for high concurrency.
Confidence: High — 3 primary sources agree
Sources:
  - prisma.io/docs/guides/performance/connection-pool (official)
  - prisma.io/blog/prisma-at-scale (official blog, 2024)
  - github.com/prisma/prisma/issues/12345 (confirmed at scale by maintainer)
Caveats: Benchmarks from 2024; Prisma 6.x characteristics may differ.

Quality checklist: 5/5 ✓
Decision: Use Prisma with PgBouncer for connection pooling.

Handing off to: /plan (architecture confirmation)
Antigravity will never present training-data knowledge as researched knowledge. If sources are inaccessible, it announces the failure reason and confidence level explicitly before presenting any findings.

/brainstorm

Run /research before /brainstorm to validate technical feasibility of approaches.

/plan

Run /research before /plan to confirm architecture decisions with sourced evidence.

/design

Run /research before /design to evaluate options with current information.

/analyze

For analysis of existing code rather than external unknowns.

Build docs developers (and LLMs) love

Get started for freeTalk to us