Put the feedback beside the work
Margin draws its findings in the margin: an underline, a circle, a bracket beside a paragraph. A mark draws attention to a passage without making the document disappear behind a report.
Opening a mark reveals the check behind it. Margin makes it possible for me to carry out a review of the writing by placing the check alongside the passage that it refers to. The full record keeps the questions, typed answers and timing available when I need to investigate further.
Fixes follow the same idea. Code proposes a few edits. Jev ranks them against the rule and checks whether they keep the point. I can compare the preferred wording with the original before deciding what to apply.
Those details serve a practical purpose. Reviewing becomes easier when the feedback stays connected to the thing being reviewed.
Each check draws its own mark
Margin ships with a small set of writing checks. Each asks one narrow question and draws a different mark, so I can tell them apart before reading the label.
Hedging gets a wave. Wordiness gets an underline. A fancy word gets a circle. Stock phrasing gets a zigzag. A paragraph that repeats itself gets a bracket, and a paragraph with nothing to flag gets a tick.
The passage below is one of Margin’s own samples. Open a mark to read its check, ask for a preferred fix and apply it.
Margin’s sample passage · preset findings
Built-in checksOur city should build more bike lanes. Cycling is cheaper than driving, it is healthier, and it takes up less road space per person.
Some people might perhaps argue that bike lanes could possibly slow traffic, in a sense. In order to be in a position to address this concern it is necessary to look at what has actually happened in cities that built them. Traffic did not get worse. Traffic stayed the same or improved.
Basically the data is clear: bike lanes work. Bike lanes are effective. They deliver results. Let’s leverage this opportunity to drive alignment on a forward-looking mobility paradigm.
Tap a mark in the passage or a note in the margin to open its check.
The passage is one of Margin’s bundled samples. Marks, colours and check descriptions come from the product; the findings and Jev preferences are preset, with no live model call.
A rule should name one concern
“Be trustworthy” is a useful aspiration. It leaves a reviewer with a lot to interpret.
Margin’s rule packs break that down into narrower questions. Did the reply imply knowledge that wasn’t available? Did it present a guess as something the person had stated? Did it take an action without clear authorisation?
The rules live in Markdown packs. Each has a name, a scope, a description and, for conversation checks, a route. That makes the expectations editable and inspectable. It also makes disagreements more useful: we can discuss the criterion as well as the result.
Authored teaching examples
Markdown rule pack## Build on something specific
scope: conversation
route: review
The reply follows up on a detail the storyteller
actually said, not a generic prompt.The reply follows up on a detail the storyteller actually shared, rather than using a generic prompt.
Nan says she made a blue dress. “What else would you like to talk about?”
Nan says she made a blue dress. “What did you like about making it?”
A paired example makes the criterion discussable. These are authored examples, not archived evaluation results.
Give each part the job it can do
Code measures things we can count, such as repeated words or long sentences. Jev supplies bounded judgements for interpretive checks. Code then applies the thresholds and routes the findings.
For conversations, the route can be auto-close, learn, review or priority review. The record shows the rule that caused it. A result near a threshold can still go to learn, giving uncertainty somewhere useful to land.
Jev returns typed answers. The explanations in the interface are assembled from those answers, the criteria and the evidence. They aren’t a transcript of the model’s reasoning.
Illustrative scores · fixed examples
Deterministic routeMove the threshold to see the local route change. Scores are illustrative, not accuracy or safety confidence.
Overall route
priority review
priority rule meets the threshold
priority reviewreviewlearnauto-closeUnapproved sharing action
priority rule · score 88
Unsupported certainty
review rule · score 72
Long reply
learning rule · score 57
Grounded, authorised reply
no review rule · score 18
The code selects the highest-priority matching route; a near-threshold priority or review result can go to learn.
Point it at any page
The Chrome extension takes Margin to wherever the writing already lives. I open it on a page, hover to highlight a section and click to pick it. The arrow keys widen or narrow the selection before anything is evaluated.
Margin then draws its marks over that part of the page itself, with the same checks, rule packs and finding cards as the app. Nothing is copied into another tool first.
In my own work it is most useful on GenAI chat responses. When the selection is a conversation, the extension reads the layout of the chat bubbles to tell the person’s turns from the assistant’s, so conversation rules can ask whether a reply answered the question, overstated its certainty or acted without permission.
The rules are what make it general. Swap in a different rule pack and the same picker can review help articles, product copy, job ads or a draft in any web editor: whatever the rules describe, Margin looks for.
Margin is invite-only while it is in beta. I’m happy to share an access code on request.
Margin is available on the Chrome Web Store.Add it to Chrome.A finding begins a review
The model sees the text supplied to it. It can assess a visible claim about an action; it cannot establish what a hidden tool actually did. It can highlight a concern; the person reviewing it still needs to consider context.
That limit affects how I author rules. I need examples of what should fire and what should pass. I need to notice false flags as well as missed problems. A new rule pack needs calibration before its thresholds deserve much confidence.
Margin is a personal tool and an evolving evaluation practice. I haven’t measured organisation-wide adoption or impact. What it has given me is a concrete way to discuss quality: here is the expectation, here is the passage, and here is the decision still requiring judgement.
Illustrative findings
Finding evidenceConstructed exchange
Nan: “I made the dress myself, from scratch.”
Choose a finding to bring its passage into view.
Review · illustrative
Grounded follow-up
Criterion
Build on a detail the storyteller actually shared.
Evidence
The question returns to the dress, a detail in the exchange.
A relevant follow-up can still be reviewed for tone and timing.