What We Changed After Reading a Thousand Advisor Edits

The edit is the real feedback form

Every advisor who uses Marlow to build a client itinerary eventually does the same thing: reads the draft, nods at the parts that are right, and then starts changing things. A hotel gets swapped. A travel day gets split in two. A layover gets flagged as too tight. None of those edits are complaints, exactly. They are corrections, and corrections carry more information than almost anything else an advisor could tell us.

A support ticket says something is wrong. An edit shows what "wrong" looked like and what "right" looks like instead, side by side, in the same itinerary, from someone who was already deep enough into the plan to know the difference. That is a richer signal than a rating or a comment box, and it is the signal we pay closest attention to as we try to make Marlow's first drafts need less fixing.

This piece is about the principle behind that process, not a tally of specific changes on a specific date. We do not think a single running count of "edits reviewed" is a meaningful thing to publish, because the value isn't in the number, it's in the pattern. Advisors correct the same handful of categories over and over, and those categories are exactly where a drafting model earns or loses trust.

Why edits matter more than ratings

Ask an advisor to rate a draft on a five point scale and you'll learn whether they were roughly happy with it. Watch what they actually change and you learn where the model's judgment diverges from theirs, and usually why.

A rating is a verdict. An edit is a diff. Diffs are legible in a way verdicts are not: you can see what was removed, what replaced it, and infer the reasoning an experienced advisor applied that the draft did not. When that inference repeats across many advisors and many trips, it stops being a one-off preference and starts looking like a gap in how the model reasons about a specific kind of decision.

That is the premise behind treating advisor edits as a design input rather than just a support queue. The edits an advisor makes on a Tuscany itinerary and the edits a different advisor makes on a Japan itinerary might look unrelated on the surface, but if both involve tightening a connection window or swapping a hotel that reads well but sits in the wrong part of town, they are teaching the same lesson twice.

The categories advisors correct most often

In principle, most of the corrections advisors make to a Marlow draft fall into a short list of recurring categories, not a long tail of unique complaints. That in itself is useful: it means the underlying causes are addressable, because they are patterns rather than noise.

Pacing is probably the most common one. A draft can be geographically sensible and still exhausting, because it treats travel days and rest days as interchangeable. Advisors who have walked a client through a trip before know that three cathedral towns in three days is a different experience than the map suggests, and they slow the itinerary down accordingly.

Hotel fit is the second recurring category, and it is rarely about star rating. A four-star property can be objectively excellent and still wrong for a specific client, because it's a business hotel on a leisure trip, or it's twenty minutes from the walkable core a couple wanted, or it caters to large tour groups when the client asked for something quiet. Advisors correct for fit, not quality, and that distinction matters.

Connection risk is the third, and it's the most mechanical of the three. A layover that is technically legal by airline rules can still be too tight once you account for a large airport, a terminal change, or a client traveling with young kids and extra bags. Advisors who have had a client miss a connection once tend to build in more margin than a schedule-optimizing draft naturally would.

Beyond those three, advisors also correct for local logistics that don't show up cleanly in structured data, such as a road that is only usable in daylight, a ferry that runs a reduced schedule off-season, or a neighborhood that is lovely at dinner but unpleasant to reach a hotel from late at night. These are judgment calls that come from having actually sent people somewhere before.

From correction to draft, in principle

None of this is useful unless it changes what the next draft looks like. The general mechanism is straightforward, even without attaching specific dates or version numbers to it.

When a category of edit shows up repeatedly across different advisors and different trips, it stops being treated as an isolated client preference and starts being treated as a gap in how the drafting process reasons about that category. A recurring pacing correction on multi-stop European itineraries, for instance, suggests the model should weight travel fatigue more heavily when a trip strings together several short stays, not just when it produces a numerically long travel day. A recurring hotel-fit correction suggests the process needs better signals about a property's character and location, not just its amenities and rating. A recurring connection-risk correction suggests the buffer logic needs to account for airport size and traveler profile, not just posted minimum connection times.

The table below sketches the general shape of this feedback loop: a category of edit, what it typically reveals, and how that lesson could, in principle, show up in a better first draft.

The result, over time, is not a single dramatic rewrite. It's a slow narrowing of the gap between what a first draft proposes and what an experienced advisor would have proposed without needing to touch it. That gap will never close completely, because taste and local judgment are moving targets, and because every advisor brings something different to a plan. The goal isn't a draft nobody ever edits. It's a draft that needs fewer, smaller, more idiosyncratic edits instead of the same predictable corrections showing up on every itinerary.

A few honest limits

It would be easy to overstate this. A few things are worth saying plainly.

• Advisor edits are a lagging signal. They tell us what needed fixing after a draft was already built, not what will need fixing on the next unusual request.
• Not every edit is a lesson. Sometimes an edit reflects a client's specific quirk, not a generalizable pattern, and treating every one-off preference as a systemic gap would just make drafts blander and more generic.
• Advisors don't always agree with each other. Two experienced advisors can look at the same draft and make different corrections, because taste and client relationships differ. The patterns worth acting on are the ones that repeat across many advisors, not any single one.
• This process doesn't replace advisor judgment, and it isn't meant to. The point of a better first draft is to give an advisor a stronger starting point so their time goes toward the parts of a trip that actually need a human's judgment, not toward fixing avoidable mistakes.
• Progress here is gradual and uneven. Some categories of edit are easier to learn from than others, and there is no finish line where a draft stops needing any correction at all.

Being honest about those limits is part of why we describe this as a principle rather than a project with a deadline. Advisors will keep editing drafts as long as Marlow keeps producing them, and that is the intended, healthy state of the relationship between a drafting tool and the professionals who use it. The edits are not a sign the tool failed. They are the mechanism by which it keeps getting better suited to how advisors actually work.

Closing thought

None of this depends on advisors changing how they work. They don't need to file feedback in a special format or flag corrections as notes for a product team. The ordinary act of fixing a draft to fit a client, done the same way advisors have always reviewed and adjusted plans, is itself the input. That is by design. The goal was never to ask advisors to teach a system. It was to build a system that pays enough attention to notice what it is already being taught.