3-Round Brainstorm: Claude x Gemini (v2.1 — Two-Layer Architecture)

Setup (first run only)

Before first brainstorm, verify these prerequisites:

Find gemini.py — bundled at scripts/gemini.py relative to this SKILL.md. Locate it with:
```
find . ~/.claude/skills -name gemini.py -path '*/gemini/*' 2>/dev/null | head -1
```
Store the result as GEMINI variable for the session.
Check GOOGLE_API_KEY — must be set. Check with echo $GOOGLE_API_KEY. If empty, look for .env file and source it, or tell the user to get a key at https://aistudio.google.com
Check google-genai package — run python3 -c "from google import genai; print('OK')". If missing, run pip install google-genai.

If any check fails, tell the user what's missing and how to fix it. Do NOT proceed without all three.

Two-layer model architecture: Flash researches, Pro reasons. Claude orchestrates both.

                  Claude (orchestrator)
                     |         |
        Phase 0.5    |         |    R1 / R2 / R3
        Phase 3.5    |         |
                     v         v
              Flash-Lite --grounded    Pro (no grounding)
              (research layer)    (reasoning layer)
                     |                  ^
                     +--Verified Context-+

Flash-Lite (gemini-3.1-flash-lite-preview, $0.25/$1.50 per 1M in/out) searches the web and produces verified facts. Pro (gemini-3.1-pro-preview, $2/$12 per 1M in/out) reasons on those facts without wasting tokens on search. Claude runs its own WebSearch in parallel during Phase 0.5 and Phase 3.5.

Version History

v1: 3 rounds, no web access. Result: 2/6 decisions invalidated by stale knowledge (A488).
v2: Added --grounded to all Pro calls. Problem: Pro wastes thinking tokens on search processing.
v2.1: Two-layer split. Flash does research, Pro focuses on reasoning. Faster, cheaper, better quality.

Why two layers beat one

In v2, Pro with --grounded does both searching AND thinking. But search processing is low-value work — it doesn't need deep reasoning to check "what version is Next.js?" Pro's thinking tokens are expensive and should be spent on challenging ideas, finding blind spots, and synthesizing arguments — not on parsing search results.

Splitting the layers means:

Flash-Lite (grounded): $0.25/$1.50 per 1M in/out, 3-5 sec per query, returns verified facts
Pro (ungrounded): all thinking tokens go to critical analysis, gets pre-verified context

Process

Phase 0: Gather Context

Before writing Round 1, check if the user provided enough context. A good R1 prompt needs:

What exists (product, tech, assets)
Hard constraints (time, team size, budget, regulatory)
The goal (what decision needs to be made)

If any of these are missing, ask the user before starting. A brainstorm with vague input produces vague output.

Phase 0.5: GROUNDING (dual research — Claude + Flash)

This phase is mandatory. Extract all technology names, frameworks, APIs, and services from the context.

Two parallel research tracks:

Claude runs 5-7 WebSearch queries (versions, pricing, compatibility, licenses)
Flash (grounded) gets a batch research prompt covering the same technologies

Run both in parallel. Merge results into a "Verified Context" block.

Flash research prompt (saved to /tmp/brainstorm-{agent}-ground.txt):

Research the following technologies and provide CURRENT (2026) facts for each.
For each one, include: current version, pricing/free tier, license, known issues.
Use Google Search to verify — do NOT rely on training data.

Technologies to research:
1. [Technology A]
2. [Technology B]
3. [Technology C]
...

Output format per technology:
- Name: current version, release date
- Pricing: free tier details, paid plans
- License: open-source? which license?
- Compatibility: works with [other stack items]?
- Red flags: any known issues, deprecations, EOL dates

Flash call:

python3 $GEMINI ask @/tmp/brainstorm-${AGENT}-ground.txt \
  -m gemini-3.1-flash-lite-preview --grounded --save /tmp/brainstorm-${AGENT}-ground-response.md

Merge: Claude reads Flash's response + own WebSearch results, resolves conflicts (if any), produces final Verified Context block:

## Verified Context (dual-checked: Claude WebSearch + Flash Google Search)
- Next.js: v16.1.6 (Oct 2025 release, 16.2 canary). Source: nextjs.org, confirmed by Flash
- Clerk: Free 10K MAU. Pro $25/mo. @clerk/nextjs@6.36.7. Source: clerk.com
- Drizzle: v0.45.x, zero-codegen, Neon HTTP driver. Source: drizzle.team
- CONFLICT: [if Claude and Flash disagree, note both and flag for R1]

Phase 1: DIVERGE (Round 1)

Claude writes a prompt with:

Verified Context from Phase 0.5 (at the top)
Full context (what exists, constraints, goals)
Initial honest assessment (strengths, weaknesses)
Instructions for Gemini to challenge, not agree

Pro is called WITHOUT --grounded — it receives pre-verified facts and focuses entirely on critical thinking.

R1 prompt template:

You are acting as an adversarial brainstorming partner. I am the lead developer framing the problem.
This is Round 1 of a 3-round brainstorm.
Your job is NOT to be agreeable -- challenge, flip assumptions, add angles I haven't considered.

IMPORTANT: The "Verified Context" section below has been web-checked moments ago.
Treat these facts as ground truth. Focus your energy on STRATEGY and ARCHITECTURE,
not on verifying versions or prices — that's already done.
If you want to recommend a technology NOT in the verified list, flag it clearly
so we can verify it before Round 2.

## Verified Context (web-checked)
[Include Phase 0.5 merged findings]

## Context
[Describe the idea, what you have, constraints, goals]

## My Initial Assessment
[Your honest evaluation: strengths, weaknesses, open questions]

## What I want from you (Round 1)
1. Challenge my framing -- what am I NOT seeing?
2. Propose 3-5 alternative angles/applications
3. Kill at least 1 idea I proposed (with reasoning)
4. Add 1 wildcard idea I haven't considered
5. If you recommend a NEW technology not in Verified Context, flag it for verification

After R1: If Gemini recommended new technologies not in Verified Context, run a quick Flash grounded check on those before proceeding to R2. Tell the user what new angles Gemini generated.

Mid-round verification (only if R1 introduced new technologies):

python3 $GEMINI ask \
  "Verify these technologies: [new tech from R1]. Current version, pricing, license, compatibility with [our stack]." \
  -m gemini-3.1-flash-lite-preview --grounded --save /tmp/brainstorm-${AGENT}-r1-verify.md

Phase 2: DEEPEN (Round 2)

Claude reads R1 response, then writes R2 prompt that:

Summarizes Gemini's ideas (including any mid-round verification results)
Kills ideas with concrete arguments (synthesize, do NOT copy-paste raw R1 output)
Adds real-world constraints (time, solo dev, regulatory, etc.)

Pro is called WITHOUT --grounded.

R2 prompt template:

This is Round 2 of our brainstorm. Here's what you said in Round 1, followed by my critical evaluation.
Your job now: STRESS-TEST my pushback, pick the 2 strongest surviving threads, and kill everything else.

All technology claims have been web-verified. Focus on architecture and strategy.

## Your Round 1 Key Ideas (summary)
[Summarize Gemini's R1 ideas — synthesize, don't copy-paste]

## My Critical Evaluation
[For each idea: KILL or KEEP with concrete reasoning]

## Constraints you must respect
[Time, resources, regulatory, skill gaps]

## What I want from you (Round 2)
1. Where am I wrong in my kills? You get ONE sentence per killed idea to defend it.
2. Pick the 2 strongest surviving threads (from ideas I kept).
3. For each survivor: concrete 2-week plan.
4. Propose 1 new idea that combines surviving thread

brainstorm