# Team Think: A Two-Agent Collaborative LLM Planning Workflow with Code Web Chat

[illumineawake]

Team Think — Two-Agent Collaborative Workflow

Recently, I've been noticing how much of a difference a solid, structured plan makes to the success of any technical implementation. A carefully created plan can significantly boost the quality of the end result, especially when dealing with complex coding tasks.

Additionally, I found great results from iterating between two different LLMs e.g. Gemini 2.5 Pro can create initial detailed plans due to its large context window, whereas ChatGPT O3 is an amazing reviewer, offering insightful critiques and actionable suggestions.

However, I started to run into practical issues:

• Large context constraints: My initial prompts became too large for O3, causing input limitations.
• Formatting inconsistencies: Different LLMs had differing interpretations of what constitutes a "plan," leading to fragmented results.
• Cumbersome workflow. Annoying manual workflow going back and forth between 2 different AIs

To resolve these problems, I formalised this workflow and optimised it specifically for managing via Code Web Chat using presets.

The result is what I call "Team Think"

Disclaimer: I am not saying this concept is anything completely original, but I think the workflow is solid and CWC definitely makes this a bit more manageable.

---

The concept behind "Team Think"

Team Think leverages two distinct but complementary AI agents, each with clearly defined roles and responsibilities:

AI 1 – Large Context Strategic Planner

This LLM should have a large context window (e.g., Gemini 2.5 Pro). AI 1’s main tasks include:

• Intake and analyze the complete initial user request and the full set of provided context files (often large, numerous, or complex).
• Create a structured plan following a predefined template, clearly outlining the background, goals, core problems, and a detailed granular checklist.
• Identify and distill context files down to exactly those required for subsequent reviews, significantly reducing the context size needed by AI 2.

This reduces the burden on AI 2, allowing it to focus purely on high-quality critique and suggestions rather than being overwhelmed by unnecessary details.

AI 2 - Specialist Reviewer

This agent is generally powered by a more focused, reasoning-optimized LLM such as GPT-4o (O3). Its primary role is to critique and refine AI 1’s structured plan by:

• Identifying gaps, edge-cases, and logical errors specifically related to coding and implementation details.
• Suggesting actionable improvements, clearly numbered and structured, making it easy for AI 1 to address feedback systematically.
• Providing deeper insights and reasoning for each identified issue, ensuring that every feedback item comes with context-rich justification.

---

How do I use it? (The Practical stuff!)

You have 2 prompts, 1 for each AI. I'll put both here and then walk through how I set this up in CWC

Prompt for AI 1 – Large Context Strategic Planner

Copy it exactly as it is from here:
https://pastebin.com/YVLdU2jE

1. You do not need to change anything with this prompt, CWC will fill in the blanks for you (next section of this guide).
2. It is explaining the Team Think workflow to AI Agent 1
3. It is specifying the exact format that it should output. I've worked through this a bit and no longer run into any formatting issues with either agent
4. It explains how feedback will be populated by AI 2 and how AI 1 should manage it. It keeps a record of the small itemised feedback, marking it as done or won't do depending on whether it gets added to the plan or not. This also prevents AI 2 repeating the same feedback suggestions if you are doing multiple iterations.
5. deletes the large feedback descriptions after absorbing it.
6. It makes a decision on whether feedback is valid/should be implemented.
7. It determines out of all of the files you have provided, what files are actually useful to your request and what files AI 2 will need to review the plan. THIS PART IS KEY!

This prompt was designed off of Google's Prompt Engineering guide they released earlier this year. There's a lot more that I could say about what is going on here, but this guide is getting very long as it is. Hopefully reading through the prompt is self-explanatory, but I'm happy to answer any questions

Prompt for AI 2 - Specialist Reviewer

Again, copy it exactly as it is from here:
https://pastebin.com/Lsy88DKp

1. You do not need to change anything with this prompt, CWC will fill in the blanks for you (next section of this guide).
2. It instructs AI 2 on how to understand the Plan format and that it should not make any direct changes to the Plan
3. Specifies how to provide feedback. It will provide small itemised feedback with a checkbox for AI 1 to mark off as complete or won't do.
4. It then provides an in depth explanation in a separate section linked to the feedback item number. AI 1 will absorb this to help with understanding the feedback and make an assessment on whether it's valid. AI 1 will then clear this section completely. We don't need it, all history is managed using the small itemised feedback.

Setting up CWC presets to streamline this workflow

Prompt 1 - Large Context AI 1

1. Create a new Preset in CWC, call it something like Team Think - Agent 1
2. Select a Large Context LLM, Google Gemini 2.5 Pro in AI Studio is the obvious choice
3. Personally I clear System Instructions and have Temperate set to 0
4. COPY everything from the prompt 1 pastebin above EXCEPT FOR THE LAST 2 LINES WHERE IT SAYS [PROVIDED FILES] and paste it in Prompt Prefix
5. Take the last 2 lines (the bit that says [PROVIDED FILES]: and onwards) and put that in your Prompt Suffix
6. Now save your preset

Prompt 2 - Specialist Reviewer

1. Follow the exact same steps as Prompt 1, but obviously use the prompt 2 pastebin.
2. Set your AI to the one you want to critique the plan. Personally I'm a big fan of ChatGPT O3 for this.
3. Save it

Now you're all setup! It should look a little like this:

AND

The Iterative Workflow with CWC Presets

Using CWC presets

Whatever you want to send to the relevant LLM will be put in by you in the Web Chat section of CWC. Because of how you have setup your presets this will put your instruction in the right spot of the prompt, alongside whatever files you pick.

Starting out a new plan would look like:

1. In the CWC Web Chat box put in whatever you need to detail your project.
2. Select all your files. Because this is going to the large context LLM I err on the side of providing more, as opposed to trying to leave some files out it might need.
3. Click on your Agent 1 preset and let Agent 1 do it's thing!!
4. You should now have an initial plan wrapped in markdown from AI 1. Copy it.
5. Go back to CWC and now remove anything in the CWC Web Chat section and paste the plan you just copied in there exactly as is.
6. In the <REQUIRED_CONTEXT> section of the plan will be a list of files that AI 1 has determined are necessary. Untick any other files that aren't these. THIS IS KEY TO BEING ABLE TO EFFECTIVELY WORK WITH AI 2.
7. Press on Preset 2 for Agent 2 (the reviewer).
8. Now watch the reviewer critique the plan, add itemised feedback and detailed feedback below your plan (which should be untouched), all wrapped up in a nice markdown. COPY IT.
9. Paste that into CWC Web chat and click preset 1 again (for Large context LLM)
10. Now this time the LLM will take the feedback, analyse it and add it to the plan if valid. It will mark off the itemised feedback (with done or won't do) and completely clear the detailed feedback.
11. Now you'll have an updated plan that has completed 1 full iteration. And the beauty of it? You can just keep repeating that process now until feel like you are no longer getting any useful feedback. This is a good indicator that your plan is ready!

What's great is your plan is fully structured in a way that another AI LLM that does the actual implementation can easily follow and mark off items. This should be straight forward for you to setup (again use templates to save time). But this post is way too long so I won't cover it here.

Lastly

If anyone actually bothers to read this and has some insight or knowledge they think could improve this further I'm all ears. I'd love to automate this even more, doing that without interfering with TOS is a challenge. I think if there was a way to auto-select the suggested files from AI 1 in CWC that would make a huge difference.

[robertpiosik]

What a great write-up. Happy to see how you utilised affixes to implement this.

I think if there was a way to auto-select the suggested files from AI 1 in CWC that would make a huge difference.
#318

[robertpiosik]

@illumineawake could you reach me on telegram so we can discuss how to improve the extension for usability with such presets? @robertpiosik

[rsallar]

Hey guys, this chat is awesome. My idea is to use gemini 2.5 for AI1 and flash 2.5 for AI2. I'm not sure how the "autoselect" files in context should work. Do you have any documentation on this process?

Edit

Ok, I got it, this is the option to "autoselect" from clipboard. As a nice feature, it could be nice that the plugin would kind of detect when the prompt has a set of files and autoselect on its own.

[robertpiosik]

@rsallar yeah, I was thinking about it already. The idea is to create a new mode: API call / Distill context. Based on task description it would ask for relevant paths among files in context and use the option shown above for the response text to find them and select.

We can also enhance the yellow button "Apply response with CWC" so when a response contains some sort of list of file paths without code blocks, we could suggest context distillation.

How does it sound?

[rsallar]

The idea would be to implement this without using api calls (to avoid extra costs). My flow would be:

• Initialize chat with all code (possible with gemini 2.5 pro with large context)
• Ask chat to plan a feature and list the involved files (there should be a previous instruction to respond to this kind of questoins with a specific format)
• The "apply response" button should detect this tag and automatically select files and context and initialize a second chat (new tab?) with only the involved files.
• We keep working with the feature and refining it in subsequent calls.
• When a new feature is needed, we go back to the first chat or we just initialize a new one with all context files.

It's just an idea and It might be too tailored to my way of coding.

[robertpiosik]

What is critical is the context distillation based on listed files in the response of chat/api call. I will do it and we'll see how it goes.

[vyrti]

@illumineawake
I say about this for months, nobody listen, we must give same problem to different models for review, and then consolidate and refactor all results from all different models
Claude 4 for initial request and Gemini 2.5 Pro for refactor, are usually best bet from my small research (go for backend, typescript for frontend), so combinations will be highly depend on request context, and programming language, as they are model specific

Also Gemini is extremely good at go lang, but all models fail hard on rust
OpenAI sometimes gives something valuable, and DeepSeek R1 gives that in even rare cases

But approach to use several (like at least 2, but I think 6+ better) LLM is the future.
Especially considering that hallucination factor for latest gemini 2.5 pro went down from 15-25 to 5.
And some LLM has 1600tk/s speed ( https://inference.cerebras.ai/ ), so current 3 minutes per request for Gemini 2.5 Pro will vanish as well
So, less they hallucinate, better the overall results from combining of models.

A full research with all possible models and combinations is needed. This may lead us to ultimate vibe coding agents. With existing models.

P.S. Even if you plan is 3 pages long, Gemini 2.5 fallback to put empty code for functions and write: Will implement later, dummy template etc, so It does not implement everything in a plan, and also if context is 1kk, the output at one iteration can be only 64k as I understood.

BUT

If the plan has sections with numbers, or chapters, then gemini FINISH everything, u just need to type
finish chapter 2
finish chapter 3
in the same chat

For example if u ask gemini to write text about Tokyo for 40 a4 pages he wont, but if u ask it to create plan, devine it to chapters, and send additional 12 prompts like above then it does it without issues!

So this means that LLM that implement the plan must get not 1, but multiple promts with like
Implement Phase 1, step 1.1
Implement Phase 1, step 1.2
Implement Phase 2, step 2.1
and all of this must stay in same context, and then we can just send usual final request:
Cant type blablabla give me full files for easy copy paste
or implement other mechanism that will gather all files from context, not just latest iteration

4. Proposed Steps:
   • Phase 1 – …
     • Step 1.1 …
     • Step 1.2 …
   • Phase 2 – …
     • Step 2.1 …
     • Step 2.2 …

[decsters01]

parabens e realmente incrivel

[cmdaltctr]

Hey! We were chatting on Reddit on my Gemini MCP. I might include your Team Think in my Gemini MCP full version, as it probably uses a different approach than the MCP approach. I'm currently looking at the A2A approach. It's in progress, but I'll need to test your approach; it looks very effective.

[detroittommy879]

Yooo...

Looks like I found some people thinking similarly to myself. I made something similar to Code Web Chat but its not a VS Code extension it runs as a python Qt app aicodeprep-gui

I have been thinking about automating my common workflows.. for really hard problems, usually i go IDE ---> o3, o4-mini, Gemini 2.5 Pro, lately though Kimi K2... that model is fantastic.
I was thinking about adding a visual 'brain' builder thing to my aicodeprep-gui app where you can choose which models to send context+problem to, and it will route them all back to a big model (maybe Gemini 2.5 Pro) to analyze all the outputs. But maybe this is overkill.. but sometimes in real life I do things like that because it works.

Nothing beats those web chat interfaces. I think the future will be using more than one model as standard. Smaller agent models to do the file edits, MCP, etc and multiple big models for difficult stuff.

[robertpiosik]

Hi, cool!

Why do you feed responses from multiple models to yet another message? Isn't it enough to manually look through them or try apply them and compare results in a live app? This is what I typically do.

Smaller agent models to do the file edits

They're pretty weak at diff patches, thus even simple updates need larger model if we want to have them fast. Smaller models are fast though, so for smaller files, when used in "whole" edit format they're fine!

[detroittommy879]

Why do you feed responses from multiple models to yet another message? Isn't it enough to manually look through them or try apply them and compare results in a live app? This is what I typically do.

Yeah that is actually usually what I do too so it might not be worth bothering with, but maybe worth experimenting with 🤷‍♂️. I have tried it manually, paste many model outputs + prompt "can you check all these potential solutions and come up with a 'best of all' best solution?" But I haven't done that enough to know if it's worth doing often.

Smaller agent models to do the file edits
They're pretty weak at diff patches, thus even simple updates need larger model if we want to have them fast. Smaller models are fast though, so for smaller files, when used in "whole" edit format they're fine!

Oh which models? I guess GPT 4.1 isn't really "small" but seems really good in Cline or any agent mode. Unlimited 4.1 for $10 w/ Copilot API. I haven't found anything smaller that is reliable like 4.1.. I haven't really experimented that much with all the newer smaller ones. But I love using the web chat models (especially Kimi K2) ---> GPT 4.1 in Cline just because it's "free" to do that.

I find that GPT 4.1 can accept "nuance" input and correct some minor errors from the bigger ones. GPT 4.1 mini starts to fail a lot with edits.