Driving an AI Session on Your Resume, Turn by Turn

Plenty of guidance exists on what to ask an AI tool about your resume. Much less exists on how to run the conversation, which is where most sessions actually go wrong. The failures are structural: one chat asked to do six jobs, a revision loop that degrades instead of improving, formatting decisions made silently, and a copy-out step nobody plans for. None of them are about prompt wording.

One task per session

Open a separate chat for each job: one for the experience section, one for the summary, one for the mechanical consistency check. Do not run them in sequence in a single conversation.

The reason is that a chat carries everything said earlier as context, and earlier turns influence later ones. If you spent twenty turns polishing your summary into something punchy, the bullets you draft afterwards in the same window will come back punchy too — the register carries over, and register in resume prose usually means inflation. Worse, if an invented detail entered the conversation at turn four and you rejected it, it is still in the transcript, and it can reappear at turn twenty in a different sentence.

A fresh session has only what you paste into it. That is the property you want.

Decide what you are pasting before you paste it

Everything you paste into a chat tool leaves your machine. Before you paste a whole resume, take thirty seconds over what is on it:

Contact details, address, and date of birth serve no purpose in a drafting conversation. Strip them.
Employer names are usually fine and often necessary for context, but if you are working on a role you have not told your current employer about, consider using a placeholder.
Anything under a confidentiality obligation — client names, unreleased product names, internal figures — should be generalised before it goes in, not after. A model cannot un-see it.

You are not being paranoid; you are deciding deliberately rather than by default. And the generalising is work you would have to do anyway, because a resume readable by an outside employer cannot contain internal-only names.

The turn order that works

Run the conversation in this sequence. Each turn hands the model something you produced.

Turn 1 — supply, do not ask. Paste your raw material for one section, from your facts file, plus the constraint. Something like:

Below are my rough notes on one job, written badly on purpose. Turn them into four resume bullets. Use only facts present in my notes. Do not add numbers, scope, seniority, or outcomes I have not written. If a bullet seems weak, say so rather than strengthening it.

That last clause matters more than the rest. Without it, “weak” gets solved by addition.

Turn 2 — ask what is missing, not for more. What would a reader want to know that my notes do not say? You get a list of gaps. You fill them from memory. The model never becomes the source.

Turn 3 — one specific complaint. Not “make it better.” Bullet two is vague about what I actually did — rewrite it using only my note about the supplier scheduling. Named target, named material, named constraint.

Turn 4 — stop. More on this below.

Why revision four is worse than revision two

Iterative revision on prose has a shape worth knowing: the first pass usually improves things, the second sometimes does, and by the fourth the text is moving without getting better. Ask repeatedly for a tighter, stronger, more impactful version of the same bullet and you will watch specificity drain out of it. “Rebuilt the weekly close process for three regional teams” becomes “streamlined cross-functional reporting operations,” which is shorter, sounds more senior, and says nothing.

This happens because each revision optimises for the instruction rather than the meaning, and the concrete detail is what an averaging process removes first. The detail was the reason the bullet worked.

So: keep the earliest version you liked. Copy it out of the chat into your own file before asking for another one. When revision four arrives, compare it to what you saved rather than to revision three — comparing to the immediately previous version hides the drift, because each step is small. Comparing to the original makes it obvious.

If two revisions have not fixed a line, the problem is not the phrasing. It is that the underlying material is thin, and no amount of rewording adds information. Go back to your notes.

When to restart rather than repair

Abandon the session and open a fresh one when:

The model has started apologising. Sequences of “you’re right, let me correct that” tend to produce increasingly hedged, increasingly generic output.
You have corrected the same invention twice. It is in the context now. A new window is faster than arguing.
You have lost track of which version is current. If you cannot say what changed since the last thing you saved, stop and rebuild from your own file.
The output has stopped containing your words. Skim for phrases you actually wrote. If none survive, the conversation has drifted off your material entirely.

Restarting feels wasteful and is not. You paste your saved good version plus one instruction, and you are further ahead in two turns than in twenty.

The copy-out step

The session ends when text leaves the chat window, and that transfer is a real step with real failure modes. Copy into a plain-text editor first, never straight into your laid-out resume. What you are stripping:

Formatting markers that arrive as literal characters.
Curly quotes and long dashes that some application forms mangle.
Line breaks inserted for on-screen reading that become paragraph breaks in a document.
Any date the model reformatted while rewriting around it. Dates are the field most likely to be checked against a third party’s records — resume dates that have to match explains why that is the one to re-type by hand.

Then read every line against your notes before it goes into the document. That is the verification pass, and it has its own method in how to check an AI-written bullet against your own record. Doing it at copy-out time rather than at the end means you are reading four lines with attention instead of a whole page with fatigue.

What the whole session cannot fix

A well-run conversation still cannot decide what belongs on the page, cannot recall what you did, and cannot judge whether this emphasis suits this employer. Those limits do not move with better session hygiene — they are covered in can ChatGPT write a resume and the field-by-field boundary is in which resume fields AI can touch.

What good session mechanics buy you is narrower and worth having: the output stays anchored to your material, you notice drift while it is still small, and you finish with a file you wrote rather than a transcript you have to excavate. If you are specifically asking for a rewrite of something you already have, the granularity of the request matters too — that is how to ask for a resume rewrite.