Writing Probe
Most writers can't see the gaps in their own reasoning — they already know what they meant to say.
Writing Probe is a Chrome extension that reads what you draft in Google Docs and surfaces sharp, Socratic questions about your reasoning, then helps you revise. No API key, no account, no setup.
TL;DR
Writers can't see the gaps in their own reasoning, because they already know what they meant to say. The sentence on the page reads as complete to the person who wrote it, even when the argument underneath it doesn't hold. Grammar checkers don't fix this — they optimize the surface, like style and grammar. None of them question what is being written, whether a claim is supported, or whether an assumption was ever examined.
Problem
Existing tools improve how text is written. None test what it argues.
Who it's for
Anyone writing serious prose where the quality of thinking matters — students on essays, researchers on papers, founders on strategy docs.
Constraint
A browser extension can't interrupt drafting flow, and it can't ask for setup before it delivers value.
I worked backward from the end behavior I wanted: a writer pauses, reads a sharp question, thinks harder, writes better. Every design and engineering decision follows from making that loop frictionless. Three insights shaped the product.
The Socratic form is the right one.
Generic feedback like “add more evidence” or “be clearer” is easy to ignore. A specific question aimed at a specific passage is much harder to wave away, because it forces engagement.
Timing is the core product decision.
A question that appears too early interrupts flow; one that appears too late arrives after the writer has moved on. Getting the pause and word thresholds right is what makes the loop feel like a writing coach rather than an alarm.
One question at a time is wrong.
Any 300-word window of real writing has multiple arguable passages. A sidebar that surfaces up to six questions lets the writer work through a set at their own pace instead of waiting for the next analysis pass.
The core loop fits inside a normal writing session: write, pause, questions appear, respond, get a suggested revision, decide. No tab-switching, no copy-pasting into a chatbot, no prompting. The extension watches the document, runs its own analysis, and surfaces questions into a sidebar pinned to the right edge of Google Docs, so it never covers the writing surface.
Amber inline underlines persist while a card is active, anchoring each question to the passage it came from. Without that visible link, the feedback reads as abstract and gets ignored.
How it works
You write. Open any Google Doc and start drafting. Writing Probe runs quietly in the background.
You pause. After a short pause, and only once there's enough text for context, it analyzes your recent writing.
Questions appear. A sidebar opens with short, direct probe questions. Each one targets a specific passage where your reasoning may need clarification, evidence, or refinement.
Passages are highlighted. The relevant passages are marked with a subtle amber underline, so you know exactly what each question refers to.
You respond. Type your answer in the sidebar — a sentence or a paragraph, whatever helps you think through the issue.
A revision is suggested. It uses your answer to generate a revised version of the highlighted passage that incorporates your clarified thinking.
You decide. Click Copy & Paste to move the revision into your document, or dismiss it and revise the passage yourself.
What the questions feel like
Short, direct, and designed to feel less like a classroom prompt and more like a smart reader pushing back.
“You say this happens often — what evidence shows it is common?”
“What does better mean here: better for whom, and by what measure?”
“This assumes readers already accept your premise. What if they don't?”
“You've made three claims in a row. Which one actually matters most?”
Under the hood
Smart triggering
Waits for a 2.5-second pause, requires at least 25 new words since the last call, and stays dormant until the doc has 50+ words of context. Quiet while you write, present when you stop to think.
Inline passage highlighting
Each question is tied to its passage with a light amber underline. Only one set of highlights is active at a time, so old ones clear automatically when new probes appear.
Multiple probes per pass
Each analysis pass can surface questions across several paragraphs at once, so you work through a small set at your own pace instead of one question at a time.
No re-surfacing old questions
It tracks which paragraphs it has already probed this session and won't keep asking about the same passage. Once your paragraphs are cycled through, it resets and starts fresh.
High-quality revisions
After you answer, it rewrites the highlighted passage to reflect your clarified thinking while matching your existing tone and style.
What it's not
Not a grammar checker, an AI writer, a style enforcer, or a note-taking tool. It reads what you wrote and asks the questions most likely to make the thinking stronger. Doing one thing is the point.
Probe modes
v2 lets the writer direct how questions are framed, depending on what they're working on. Pair that with zero-setup access — a Cloudflare Worker holds the key server-side, so there's a monthly usage allowance instead of an API key to paste in.
v1 shipped as a working Chrome extension, and the core loop held: pause, question, response, suggested revision. Shipping is what exposed the friction the spec had missed — and almost all of it lived in the first few minutes of a new user's session.
What v1 shipped with
A sidebar of up to six question cards per pass, amber underlines linking each question to its exact passage, a suggested revision after every response, zero setup (install and write — no account, no key to paste in), and a monthly usage allowance that resets on its own.
Biggest v1 blocker
Before the extension could read anything, first-time users had to flip on an obscure Google Docs setting buried in the menus. A one-time step, but it sat in front of the entire product and stopped nearly every new user cold.
v2 fix
v2 removes the step completely. The tool works the moment it installs, with nothing to configure.
Accept-flow friction
After accepting a suggestion, users still had to hunt through the document to find the passage they were replacing.
v2 fix
v2 scrolls straight to it and pulses the amber highlight so the eye lands on the right spot. Small change, but it closed the last gap in the loop.
Hardest constraint
Google Docs is a canvas-rendered editor, not standard HTML — it breaks DOM assumptions, Trusted Types blocks direct injection, and class names change without notice. The extension is brittle at the extraction and highlight-injection layer by nature, so the mitigation is defensive: fallback chains, extraction health checks, and explicit error states when detection fails.
Outcome
Writing Probe shipped on the Chrome Web Store with zero-setup access, a monthly usage allocation, and the full Socratic loop running inside Google Docs. v2 is built around the failures that only surfaced after v1 shipped — primarily the screen reader requirement and the missing probe history.
Reflection
The real tension in this product is between automation and interruption. An extension that fires too eagerly becomes noise, and users learn to dismiss it without reading. The triggering logic is the product's behavioral contract with the user, and getting it right meant reasoning about the rhythm of a writing session, not just the technical constraints.