After posting several articles about the impact of AI on developers and sharing resources to help mitigate some of the risks, I wanted to share a new tool I've been building and experimenting with.
I introduce you to ThinkMode!
A chrome extension that helps developers choose a thinking mode before prompting AI, and calculate the associated cognitive cost after prompting AI.
AI makes it easy to skip thinking.
This tool helps slow down and integrate thinking before and after prompting the AI.
As a browser extension it can be easily integrated into your everyday LLM chat prompting workflow.
The Idea
Before opening a prompt, ThinkMode asks you to describe what you are trying to do.
It then recommends one of 5 modes:
- Explore: understand the problem before solving it
- Challenge: pressure-test an existing plan
- Decide: compare options and tradeoffs
- Audit: review quality, correctness, tests, and edge cases
- Reflect: learn from what you just did
The goal is to match the prompt with the kind of thinking intended for the task.
The tool also includes a manual AI usage log.
After using AI, you can log how you used it.
Usage is grouped into three categories:
- Supportive: AI helps expand your thinking
- Mixed: AI saves time but may compress understanding
- Risky: AI may replace your judgment
Each log adds to a cognitive cost meter. Riskier usage fills it faster.
Cognitive cost measures tradeoffs between using AI to expand understanding vs outsourcing judgment.
When the meter is full, ThinkMode temporarily pauses supported AI chat pages for 5 minutes.
Most developer tools optimize for speed.
I wanted to experiment with a tool that introduces intentional friction in order to surface reflection.
How it works
ThinkMode is a Manifest V3 Chrome extension built with React, TypeScript, and Vite.
The architecture includes:
- a content script detects supported AI chat pages
- a floating button opens the extension
- a background service worker coordinates the side panel
- a React side panel handles the main workflow
- shared TypeScript modules handle recommendation logic and prompt generation
The extension currently supports ChatGPT, Claude, and Gemini pages.
There is no backend nor LLM API call involved.
ThinkMode does not read conversations, scrape page content, send analytics, or store data remotely.
The recommendation engine is deterministic. It uses simple keyword rules to choose a thinking mode which felt reasonable considering this is an MVP.
What I Learned
Sometimes the useful AI tool is not the one that gives you a better answer.
It can be the one that helps you ask a better question.
I'm curious:
Have you ever caught yourself accepting an AI answer before fully understanding the problem?
If so, what habits or tools help you stay engaged in the thinking process?
How to install
You can download the extension free here from the chrome web store
Otherwise you can run the extension locally using developer mode from chrome extensions. The code is open source here on github.
Feel free to fork it and adapt it to your own needs.
Would really appreciate a star or review if you find it useful!
If you can think of ways to improve this tool or want to see other features, let me know in the comments!
Top comments (30)
Cognitive cost tracking is the interesting part. Most friction tools just slow you down, but showing cumulative cost over sessions builds actual data for the habit. That feedback loop is what most think-before-prompting advice is missing.
Exactly! I think there is real value in this data over time for increased awareness and self-improvement. Maybe AI tooling can even adapt intelligently based on how high or low your congitive cost is at a certain time. For eg. become more restrictive in its answers if you have a higher cognitive cost, to push you to think more independently.
the adaptive part is the gap no one has built yet. knowing your cognitive load is one thing, actually changing how the tool responds to it - that is the harder and more interesting piece.
100% it is an interesting area to look at, feels like there is a lot that can be done to make AI tooling feel more adaptive to each user, as every user has different needs and goals
right, and the tricky part is that most tools adapt to tasks, not to the person doing the task. building per-user adaptation means modeling your state, not just your input — which is a meaningfully harder loop to close.
Good stuff Julien. Was waiting for this article for a while lol
Thanks Francis. True I did mention this project to you a few times already :p
the 10 second pause is clever but what catches my eye is the prompt challenge itself - forcing you to articulate the goal before you hit generate. we have seen similar gates in deployment pipelines: approval steps that require typing a reason actually surface bad pushes before they happen, not after. the articulation as filter pattern works because vague requests usually mean vague thinking. did you track what people typed in the challenge vs what they would have typed without it? that prompt delta seems like the real signal worth measuring
Thanks Mudassir for your inputs.
That would indeed be interesting to know, however I don't track anything with this tool, it uses your local storage.
local storage makes sense for privacy — no backend means no question about who sees the data. the tradeoff is you lose the aggregate signal: right now the only person who can see 'my prompts get sharper after 3 days of forced articulation' is the individual user, not you as a product builder. does thinkmode have any export or self review mechanic, or is the data strictly ephemeral?
it does not have any export feature yet, but that is a great suggestion Mudassir, I will definitely consider this
fair — the privacy first tradeoff makes sense for early adoption, even if you lose the population level signal.
one approach if you do add export: JSON dump of session metadata only (timestamp, cognitive cost score, mode selected) without the actual prompt text. you keep the behavioral pattern data without storing what people typed. even a simple copy to clipboard flow would let power users build their own dashboards.
are you planning to open source it at some point?
That would be a good approach indeed, thanks.
It is open source already: github.com/JulienAvezou/ai-thinkin...
The architectural read on this one clicked for me. The no backend, no LLM API call, no chat history relay posture is the actual point. It is what makes ThinkMode believable as a reflection tool rather than another tracker.
The thread Mykola started on adapting to the person rather than the task is the right next problem, and the privacy first tradeoff Mudassir flagged is real. One observation from walking a similar no backend path: the permission surface is the cheapest proxy you have for how much the user should trust the extension. ThinkMode does not need , does not need host permissions on chatgpt.com / claude.ai / gemini.google.com, and the content script runs only when the user is already inside one of those pages. That posture is what lets the cognitive cost framing land. If a user reads "we pause your AI chat for 5 minutes when the meter is full" they will only believe it if the extension cannot itself reach into those pages.
The piece that compounds on the same architecture is going from the inside AI page case (yours) to the in page question case. Same posture: no host permissions on the AI destinations, content script runs only on demand when the user triggers selection or a screenshot+annotation pass, side panel opens with a per question destination picker. Same privacy story extends because nothing is being scraped on load. The selection to side panel to AI destination path is a different user moment but it does not require a new permission model, only a new trigger.
On the aggregate signal gap, JSON session metadata export without prompt text (timestamp, cognitive cost score, mode selected) is the right shape. Users who care can pipe it into their own dashboard. You keep the behavioral pattern data, lose nothing users actually typed, and you do not become the storage layer for 10k people thinking out loud.
Julien, I love the concept behind this. Most tools try to help us move faster, so it's interesting to see one that's encouraging us to slow down and think a bit more.
Just gave it a star, and I'll be trying it out soon. Thanks for sharing it.
Thanks for the support Hemapriya! Much appreciated.
If you have any feedback after trying it, please let me know :)
Building a Chrome extension to make AI use more intentional is a great way to encourage thoughtful and productive interactions with AI tools. The extension could help users set goals before starting a conversation, track AI usage patterns, provide reminders to verify important information, and reduce over-reliance on automated responses. Features like prompt templates, focus modes, and usage analytics can help users get more value from AI while maintaining critical thinking.
From a marketing perspective, businesses can apply the same principle by using AI strategically rather than relying on it for every task. Companies like Aqva Marketing leverage AI to enhance digital marketing efforts, streamline workflows, and analyze data, while still combining human creativity, expertise, and strategy to deliver meaningful results. The most effective approach is using AI as a tool to support decision-making, not replace it.
Intentional AI usage leads to better outcomes, whether you're building software, creating content, or growing a business through smart digital marketing practices.
I like the way you framed this akash! Intentional AI makes a lot of sense.
Love the intentional design here — the "thinking mode" selector before prompting is exactly the kind of metacognitive layer missing from most AI tools.
The cognitive cost angle is especially interesting for writers. When AI can generate and publish a full blog post to 15+ platforms in one click (like our twRty Blogboat does — twrty.org/blogboat), the risk is the same: speed without intentionality kills quality.
The best workflow is probably: AI handles the mechanical friction, humans stay accountable for the thinking. ThinkMode seems like a great forcing function for that. Would love to see a writing-focused mode in there — something that asks "does this idea deserve to be published?" before generating content.
I love the writing focussed mode idea, thanks for suggesting.
I like this idea.
Most people do not misuse AI because they are reckless. They misuse it because it becomes automatic. A small reminder layer before prompting can help people slow down, avoid sensitive data sharing, and use AI with more purpose.
Exactly, thanks Suny.
Very nice!
Thanks Austine!
Thank you for this!
Thanks Jasmine! I am glad you like it.
Some comments may only be visible to logged-in visitors. Sign in to view all comments.