The AI Research Stack: Claude, Perplexity, and NotebookLM in One Workflow

Every general-purpose AI assistant is mediocre at deep research. The reason is structural: a single model can't simultaneously crawl the live web, hold a 1M-token corpus in working memory, and write a tight synthesis. Senior researchers stop trying to make one tool do all three and split the workflow across three.

This is the stack I run. Perplexity for live discovery. NotebookLM for grounded synthesis over a fixed corpus. Claude for the writeup and quality control. The handoffs between them are where the speed comes from.

Accurate as of April 2026. Claude 4.6, Perplexity Pro with Comet enabled, NotebookLM on the AI Pro tier with Gemini 2.5 Deep Research. UI and feature names move quickly — verify against the vendor pages if you're reading this months later.

Why three tools and not one

A research task has three phases that don't share gradient. Discovery (what's out there?) needs current web access and breadth. Synthesis (what does the corpus say?) needs grounding on a controlled set of sources you trust. Writing (what's the argument?) needs reasoning depth and clean prose.

ChatGPT's Deep Research mode tries to do all three in one shot. It's fine for shallow questions. For anything you would stake your name on, it hallucinates citations and conflates sources. Splitting the work fixes that, because each tool is bounded to what it is good at.

The workflow

There are four steps. Don't skip them — the handoffs are the point.

1. Frame the question in Claude. Open a Claude project. Drop in any prior writing you have on the topic, plus a one-paragraph statement of what you are trying to learn. Ask Claude to (a) restate the question more precisely, (b) list the sub-questions that must be answered, and (c) propose six to ten candidate search queries.

The output is your discovery brief. Most researchers skip this and start Googling. That is why their research drifts.

2. Run discovery in Perplexity. Paste each search query into Perplexity Pro with Comet enabled. Comet runs queries as an agent — it follows links, reads the underlying pages, and returns a synthesis with citations. For each query, save:

• The three to six most relevant URLs Perplexity surfaces
• Any direct quote you will likely cite later
• One sentence on why the source matters

Don't trust Perplexity's synthesis as your answer. It is a discovery engine. You are harvesting URLs.

3. Build the corpus in NotebookLM. Create a new notebook. Drop in every URL you saved (NotebookLM accepts up to 50 sources on the paid tier — more than you should need). Add any PDFs, internal docs, or transcripts you want grounded.

Now ask NotebookLM the sub-questions from step 1. Its answers will be cited inline against the sources you uploaded. Read them critically. NotebookLM rarely hallucinates inside the corpus, but it will under-weight a source if the wording doesn't match the question. If an answer feels thin, ask the question three different ways.

Export your findings: select the relevant chat turns and copy them out, or use Save to Note to pin the strongest answers.

4. Write in Claude. Back to the Claude project. Paste in the NotebookLM findings, the original question, and your sub-questions. Ask Claude to draft the synthesis. Then — and this is the step everyone skips — paste the draft back into NotebookLM and ask: "Are any claims in this draft unsupported by my sources?" NotebookLM will flag them with the missing citation. Fix those. Repeat until it returns a clean pass.

What this saves you

A traditional research workflow has two failure modes: shallow synthesis from one tool, or hours lost rabbit-holing on irrelevant sources. The stack above caps the rabbit hole at step 2 (Perplexity is fast enough that you bail on weak queries in seconds), grounds the synthesis at step 3 (no hallucinated stats), and audits the writeup at step 4 (no unsupported claims).

End-to-end on a moderately complex question: about an hour. For a serious deep dive: half a day instead of three.

The prompts that matter

Two prompts do most of the work. The first is the framing prompt for step 1:

You are helping me research [topic]. My goal is [specific question]. Restate the question more precisely, list four to seven sub-questions that must be answered to address it, and propose six to ten search queries that would surface the best sources. Be specific. Avoid generic phrasings.

The second is the audit prompt for step 4, run inside NotebookLM after pasting the Claude draft:

Review the following draft against my uploaded sources. List every factual claim, statistic, or attributed quote. For each, indicate whether it is supported by a source in this notebook (cite the source) or unsupported.

These two prompts are 80% of the leverage. The rest is plumbing.

Where this breaks

The workflow assumes the topic is researchable from public web sources. For internal or proprietary domains, swap Perplexity for a private search index (Glean, Coveo, or your own vector store). The rest of the pattern holds.

It also assumes you trust the sources you uploaded. NotebookLM doesn't fact-check the corpus — it grounds against it. Garbage in, grounded garbage out. Build the habit of pruning weak sources before step 3, not after.

For workflows that need to update continuously (newsletters, market monitoring), schedule the discovery step rather than running it once. The patterns in AI automation workflows cover the wiring. Pair this with advanced prompt engineering for the framing and audit prompts, and the AI-powered job search system becomes a downstream variant of the same workflow.

Cost and access

Perplexity Pro with Comet runs $20 per month. NotebookLM with Gemini 2.5 Deep Research is included in the Google AI Pro tier at $20 per month — the free tier exists but the source cap and reasoning quality are noticeably tighter. Claude Pro is $20 per month. Total: $60 per month for a research stack that replaces an analyst-hour-per-week.

If you are running this commercially, the API equivalents (Anthropic, Google AI Studio, Perplexity API) are cheaper at low volume but require you to write the orchestration yourself. For most solo operators, the consumer tiers are the right call.

Frequently Asked Questions

Can I do this with just ChatGPT's Deep Research mode?

You can, for shallow or low-stakes questions. For anything you would publish or present, the lack of an audit step is a real problem — Deep Research will assert confident claims that turn out to be paraphrases of bad sources. The three-tool split exists specifically to prevent that.

Why NotebookLM instead of just uploading sources to Claude?

Claude can take long context, but it doesn't enforce strict grounding the way NotebookLM does. NotebookLM is purpose-built to refuse claims that aren't in the corpus. That refusal is the audit. Claude is better at the writing step than at policing itself.

How many sources should the corpus have?

Five to fifteen for a typical question. More than that and the synthesis gets diluted unless the sources are tightly scoped. Fewer than five and you are probably under-researched.

Does the order of steps matter?

Yes. Frame before discovery, discover before synthesis, synthesize before writing, audit before publishing. Skipping the framing step is the most common mistake — it is where you lock in what counts as a good source.

What about Gemini, Mistral, or DeepSeek for the writing step?

Substitute freely. The pattern is tool-agnostic at step 4 — any frontier model with strong reasoning works. I default to Claude because long-context handling and willingness to push back on weak arguments matter more than raw speed at the writeup phase.