Sometimes One LLM Isn't Enough
The first time I encountered the term ADR (Architecture Decision Record), I asked Claude Code: "What's an ADR?" The response was "Architecture Decision Record. It records important decisions." Technically correct, but I learned nothing.
Claude Code is a development-focused agent. It writes code, edits files, and runs tests. But when you ask it to explain a concept in depth, it tends to give terse answers because it's operating within a development-task context.
I use a 4-stage approach with four different LLMs.
The 4-Stage Approach
Stage 1: Ask Claude Code
Start by asking Claude Code about the term or concept that came up during development.
me: "What's an ADR?"
Claude Code: "Architecture Decision Record. It records important decisions."
You get an answer scoped to the development context. For information directly tied to implementation, this is often sufficient.
Stage 2: Ask Gemini / Standard Claude
When Claude Code's answer feels insufficient, ask Gemini or Claude (the standard chat version) the same question.
me: "Explain ADR in a way a beginner would understand"
Gemini: "An ADR is a document that records the reasoning and
background behind important technical decisions in a project.
For example..."
You get more detailed explanations, concrete examples, and beginner-friendly breakdowns.
Stage 3: Ask for an Analogy
When you still don't get it, try a different angle.
me: "Explain TDD using an analogy from Baki the Grappler"
Claude (standard): "TDD is like first 'anticipating your opponent's
technique' and then 'working out a counter'..."
Mapping an abstract concept onto a domain you already know makes it concrete. The subject can be anything — manga, sports, cooking.
Stage 4: Deep Research
When you want a thorough investigation, fire the same question at multiple LLMs in parallel.
Here is how I personally split them (as of February 2026):
| LLM | Why I use it | How I use it |
|---|---|---|
| Claude Code | Practical implementation answers | "How should I use this in my project?" |
| Gemini | Detailed concept explanations | "Explain this for a beginner" |
| ChatGPT | Best practices | "What's the generally accepted approach?" |
| NotebookLM | Information synthesis | Feed multiple sources for cross-reference analysis |
Using NotebookLM
I use NotebookLM specifically for "synthesizing multiple sources."
- Copy Claude Code's answer
- Copy Gemini's answer
- Add relevant documentation URLs
- Feed everything into NotebookLM
NotebookLM cross-references the sources and organizes contradictions and common points.
For example, when researching ADR templates, Claude Code said "5-section structure" and Gemini said "7-section structure." After feeding both into NotebookLM, it clarified: "The baseline is 5 sections; 7 sections is recommended for team workflows."
Decision Flowchart
A question comes up during development
↓
Ask Claude Code
↓ Answer is sufficient → Go back to development
↓ Answer is insufficient
Ask Gemini / Claude
↓ Understood → Go back to development
↓ Still unclear
Ask for an analogy
↓ Understood → Go back to development
↓ Want to go deeper
Deep research (multiple LLMs in parallel)
Most questions are resolved at stages 1-2. I only reach stages 3-4 when learning a concept for the first time.
Takeaways
Since adopting this approach, the time I spend understanding new concepts has roughly halved. I used to push forward with a vague sense of "I sort of get it," only to hit walls later. Now I make sure I genuinely understand at stages 2-3 before returning to implementation, and rework has decreased.
- Claude Code excels at implementation-focused answers. Ask it first
- For conceptual understanding, supplement with Gemini or standard Claude
- Analogies are a breakthrough tool for understanding. Map concepts onto domains you know
- Use NotebookLM to synthesize multiple sources and resolve contradictions
- Decide in advance "which LLM to ask for what" so you don't waste time deliberating
Top comments (0)