Most AI output feels mediocre for a reason that has nothing to do with the model. You asked one prompt to brainstorm and decide at the same time, and the two jobs cancel each other out. This guide gives you the two-mode framework, the exact prompts for each mode, and a map of which tool to use when.
Here is what a normal session looks like. You open Claude Code, you start typing, and inside one prompt you are researching, generating ideas, weighing them, and asking it to build something. Everything happens in the same breath.
The output comes back thin. Not wrong, just average. You blame the tool, tweak the wording, and run it again. Same result.
The problem is not the tool. You mixed two kinds of thinking that work against each other. Generating options and filtering options are opposite motions, and doing both at once means you never fully do either. Separate them, use the right tool for each, and run them in sequence. That is the whole method, and the rest of this guide is how to run it.
The Two Modes
Every task you bring to AI wants one of two things: more options, or fewer. Those are the two modes.
Divergent thinking generates. The goal is volume: twenty ideas, fifteen angles, ten approaches. The mindset is "yes, and," where nothing is off the table yet. You reach for it when you feel stuck or blank and are not sure where to start.
Convergent thinking decides. The goal is precision: one plan, one deliverable, the single best option built out. The mindset is "yes, but," where you apply real constraints and criteria. You reach for it when you feel buried under too many options and need clarity.
| Divergent | Convergent | |
|---|---|---|
| Goal | Generate ideas, angles, possibilities | Filter, select, structure, build |
| Mindset | "Yes, and": nothing off the table | "Yes, but": constraints and criteria |
| Output | Volume: 20 ideas, 15 angles | Precision: 1 decision, 1 plan |
| Use it when | You feel stuck or blank | You feel overwhelmed by options |
| The failure mode | Filtering too early kills the ideas | Generating more when you need to pick |
| Good tools | NotebookLM, Perplexity, voice-to-text | Claude Code, structured build prompts |
The cardinal rule sits under that table: never run both at once. The moment you start judging ideas mid-brainstorm, the creative flow stops. The moment you start inventing new options mid-decision, execution slips another day. Keep the modes apart, put each in its own tool, and do them one after the other.
Rule Zero: Start With What Done Looks Like
Before you open any tool, answer one question. What does done look like?
The answer tells you which mode to start in. This is ordinary project management applied to an AI workflow, and skipping it is why most sessions wander.
Start divergent when you do not know what to build yet, when the problem itself is still fuzzy, when you need options before committing, or when nobody has defined the scope. You are working from a blank canvas.
Start convergent when the opposite is true. You already have too many ideas and need to pick one. Or the problem is clear and only the execution needs structure. Or you inherited a pile of notes, transcripts, and research that needs synthesizing before anything else can happen. You are working from a messy pile.
| Start divergent when… | Start convergent when… |
|---|---|
| You don't know what to build yet | You have too many ideas and need to pick |
| The problem space is unclear | The problem is clear, the solution needs structure |
| You need options before committing | You have notes or research to synthesize |
| You're starting from scratch | You're refining something that exists |
| Scope isn't defined | Scope is defined, the plan isn't |
The shortcut: if your goal is to discover what to do, start divergent. If your goal is to deliver something specific, start convergent. Most projects run divergent first and converge second, because you explore before you build. But when you inherit a backlog or a client brain dump, you flip it and filter first.
Pattern One: Explore, Then Build
This is the common one. Use it when you are starting fresh and need to open up the problem before you commit to anything.
Phase one, divergent: generate
Your only job here is volume. No filtering, no judging, no "but that will never work." Reach for tools built to expand.
NotebookLM works when you have source material. Upload the docs, research, or transcripts you already have and ask it for fifteen possible approaches, and the ideas come back grounded in your actual files rather than invented from nothing. Perplexity works for research-first ideation: ask what other companies are doing about a problem and it gives you current, sourced angles you would not have found alone. Voice-to-text tools like Wispr Flow or Otter work for raw thinking, where you talk freely for ten or fifteen minutes and the unedited transcript becomes your divergent input. A plain chat with Claude or ChatGPT works too, as long as you ask for breadth on purpose.
Copy the divergent prompt:
Copy this.
I'm building an automation system for a marketing agency that manages 30+ client accounts. They spend 20 hours a week on manual reporting. I need ideas, not solutions yet, just ideas. Give me 20 different angles I could explore. Include the obvious ones and the weird ones. Don't filter for feasibility.
Notice what the prompt forbids. It bars the tool from filtering, because feasibility is the next phase's job, not this one's.
Your first real run: pick a problem you are actually chewing on and ask any chat tool for twenty angles with that exact "don't filter for feasibility" instruction. Keep every idea, including the bad ones. You are collecting raw material, not answers.
The mistake that makes it fail: judging as you go. The second you think "that one is stupid" and skip writing it down, you have started converging, and the flow dies. Write down the stupid one. You can cut it in ninety seconds during the next phase.
Phase two, convergent: filter and build
Now switch tools and switch your head. Take everything from phase one and bring it to Claude Code, which is the convergent specialist. Its job is to analyze, rank, select, and build.
Copy the ranking prompt:
Copy this.
Here are 20 ideas I brainstormed for automating a marketing agency's reporting workflow. Score each one against these criteria: - Implementation complexity (1-5) - Time saved per week (estimate) - Client-facing impact (high / medium / low) - Can be built with n8n + Claude Code (yes / no) Rank them. Give me the top 5 with your reasoning.
Once you have the top five, go deep on the winner:
Copy this.
Take the #1 ranked idea and build me: 1. A system architecture, described in text 2. The implementation plan: phases, tools, timeline 3. The first working prototype. Start coding.
Why the split works: in phase one you used tools that are good at opening up, and they handed you raw material. In phase two Claude Code takes that pile and turns it into a ranked decision and a running prototype. You used each tool for the one thing it does best instead of asking a single tool to do both jobs badly.
Pattern Two: Filter, Then Expand
Use this when you already have too much. Not a blank page, an overflowing one.
Sometimes the problem is not a shortage of ideas. You have a forty-minute voice memo, thirty pages of meeting notes, a Notion database with two hundred feature requests, or a client who just talked at you for an hour. There is no lack of input. There is a lack of structure, and piling on more ideas makes it worse.
Phase one, convergent: extract and structure
Start by compressing. Feed the mess to Claude Code and ask it to pull the signal out.
Copy the transcript-extraction prompt:
Copy this.
Here's a raw transcript from a 40-minute strategy call with a client. Extract: 1. The 3-5 core problems they mentioned, with quotes 2. Their stated priorities versus their implied priorities 3. Decisions that were made versus ones still open 4. Contradictions: things they said that conflict 5. The single most important thing they need solved first
The same move works on a backlog:
Copy this.
Here are 200 feature requests from our users. Cluster them into themes. For each theme tell me: - How many requests fall into it - The core job-to-be-done behind the requests - One sentence describing the ideal solution Then rank the themes by frequency times impact. Top 5 only.
Now you have clarity instead of noise. A ranked shortlist, not a wall of input.
Phase two, divergent: expand on the winners
With the top priorities named, switch back to divergent mode and explore how to solve each one. Take a winning theme to NotebookLM, Perplexity, or a fresh brainstorm.
Copy this.
The #1 user theme is 'automated client reporting'. Give me 15 different approaches to solving this, from simple (template plus manual) to advanced (fully autonomous agent). Include approaches from other industries. Don't filter yet.
From here you loop back to convergent. Take those fifteen approaches to Claude Code, rank them, pick one, and build it. The cycle keeps going.
Real Projects Chain Both
You rarely run a single clean cycle. Real work chains several, and the skill is knowing which mode you are in right now and when to switch.
| Chain | When to use it | Example |
|---|---|---|
| D→C | Starting fresh, need ideas before building | Brainstorm 20 automation ideas, then rank and build the top 3 |
| C→D | Drowning in input, need structure first | Synthesize a 40-min call, then brainstorm fixes for the top pain |
| D→C→D | Explore, prototype, then improve it | Brainstorm features, build the MVP, brainstorm feedback angles |
| C→D→C | Filter a mess, expand on winners, then build | Cluster 200 requests, brainstorm for the top theme, build it |
| D→C→D→C | Full product cycle | Research, select, prototype, iterate, ship |
One rule holds across every chain: it must end on convergent. Divergent thinking produces possibilities, and convergent thinking is what turns a possibility into something real. Stop after a divergent phase and you own a pile of ideas and nothing shipped. The last step is always filter, decide, build, deliver.
The Tool Map
Most AI tools are good at one mode, not both. Match the tool to the mode instead of forcing your favorite to do everything.
| Tool | Best mode | What it's good at |
|---|---|---|
| NotebookLM | Divergent | Upload your materials, ask for ideas grounded in them. Ideation rooted in real research. |
| Perplexity | Divergent | Research-backed exploration with cited sources. Surfaces angles you'd miss alone. |
| Voice-to-text | Divergent | The unstructured brain dump. Talk 15 minutes, get a raw transcript, no filtering. |
| ChatGPT / Gemini | Divergent | Fast brainstorm partner. Strong on volume, weaker on hard selection. |
| Claude Code | Convergent | Analyze, rank, decide, architect, build. Takes messy input, returns structure. |
| Claude (chat) | Both | Handles either mode if you say which. Tell it "brainstorm" or "analyze and decide." |
| Cursor / Bolt | Convergent | Build mode. Executes a clear plan. Use it after the decisions are made. |
The one thing to internalize: a tool built for divergence will give you a shallow decision, and a tool built for convergence will give you a narrow brainstorm. The mismatch is where the mediocre output comes from.
A Full Project, Start to Finish
Here is how the chaining looks on one project. Picture a B2B SaaS company at three million in annual revenue. Their five-person sales team spends fifteen hours a week building proposals by hand, chasing follow-ups, and updating the CRM. They want to automate it and do not know where to start.
Phase one is divergent. You upload their current sales process doc to NotebookLM and ask for fifteen automation opportunities, small wins and ambitious ones alike. You search Perplexity for how companies are automating proposal generation. You record a twelve-minute voice memo walking through what you noticed on the client call: what frustrated them, what they repeated, where they lit up. Across the three sources you end up with roughly thirty-five raw ideas.
Phase two is convergent. You bring all thirty-five into Claude Code:
Copy this.
Here are 35 automation ideas for a B2B SaaS sales team. The client has $3M ARR, a 5-person sales team, and uses HubSpot. They care most about time saved and proposal quality. Score each idea on effort (1-5), impact (1-5), and time-to-value. Give me the top 5, ranked. For #1, outline a 3-phase plan.
The top five come back ranked. Number one is an AI-powered proposal generator wired into HubSpot.
Phase three diverges again, on a narrower scope. You ask Perplexity for the best approaches to AI proposal generation, and you upload the client's winning past proposals to NotebookLM to find the patterns worth templating. That gives you twelve approaches plus the template patterns.
Phase four converges to a build:
Copy this.
Here are 12 approaches to building an AI proposal generator and the template patterns from winning proposals. Pick the approach that works with HubSpot plus the Claude API. Build me: 1. The data flow: HubSpot deal to AI to proposal PDF 2. The prompt that generates the proposal 3. A working prototype I can demo to the client
Out comes a working prototype you can put in front of the client.
Watch the rhythm. Thirty-five ideas compress to five, five expand into twelve approaches for the winner, twelve compress into one built prototype. Each cycle narrows the scope and deepens the quality, so by the time you are writing code, every decision behind it was informed rather than guessed.
When This Framework Slows You Down
This is a structure, and structure has a cost. Worth naming the cases where it is the wrong tool.
For a tiny task, skip it. If you need a two-line email or a quick regex, do not stage a divergent phase and a convergent phase. The overhead is bigger than the job. This framework earns its keep on projects with real ambiguity, where a wrong early decision is expensive to unwind.
The tool boundaries are guidelines, not laws. Claude chat can brainstorm well when you tell it to, and Perplexity can help you narrow down. The point is not which logo you use. The point is that you decided, before you started typing, which mode you were in, so you are not silently doing both.
And the framework does not decide for you. It organizes the thinking; the judgment about which idea is actually good stays yours. If you feed it a weak brainstorm, you get a well-ranked list of weak ideas. Garbage in still ranks the garbage.
If You Do One Thing This Week
Take the next real project on your desk and answer one question before you open a tool. What does done look like. Then decide whether you need more options or fewer, and open the matching tool for that mode alone.
That single habit, naming the mode before you type, is most of the gain here. It costs you about thirty seconds of thought at the start of a session. If it does not change your output, you lost thirty seconds. If it does, you stop shipping average work from tools that were never the problem.