Apprendre votre style d'écriture à une campagne

Open the campaign, click the AI Instructions tab, then click 2 Teach the AI. The Learned writing profile card sits at the top of the left column, above the testimonials card and the explicit instruction fields.

Its description is precise about what it does and does not do. You add outbound emails you consider strong, and FirstSales extracts reusable patterns for this campaign. It does not train or fine-tune a model. The patterns stay attached to this campaign rather than becoming a workspace-wide style.

An Analyze more button sits at the top right of the card. That is how you add samples after the first pass, and it is the same entry point whether the card is empty or already active.

The card is one of two style inputs on this tab. The other is the WRITING STYLE field in the explicit instructions section below it. They are deliberately separate, and the explicit instructions section says so directly: the approved writing profile stays a separate learned input, so neither source silently overwrites the other. Use the profile for tone that is easier to show than to describe, and the field for rules you can state plainly.

If your campaign has no profile yet, drafts still work. The profile is an addition to the brief, not a prerequisite for it.

Click Analyze more and add outbound emails you rate highly. The useful sample is a real email you sent that got a real reply, not a template you like the look of. The profile extracts patterns from what you give it, so a polished sample nobody ever answered teaches a style that does not work.

Aim for samples that differ from each other in subject but match in voice. Three emails to three different industries written in your voice teach the voice. Three near-identical emails to the same company teach one email.

Avoid sending in anything you would not want reflected in a draft. If a sample contains a claim you can no longer make, a price that changed, or a customer name you cannot use publicly, the patterns pulled from it will carry that flavour into new drafts. Rewrite the sample before adding it rather than adding it and correcting the output later.

The number of samples that helps is small. A handful of genuinely strong emails beats twenty average ones, because averaging across mediocre samples produces mediocre patterns. Start with three or four, look at what the profile reports, and add more only if the extracted style is thinner than you expected.

Samples are added per campaign. A profile built for a founder-led campaign will not appear on a campaign aimed at enterprise procurement, which is usually what you want.

Once the card is active it shows a status line, "This campaign now uses the approved profile for new AI Autopilot drafts", and three counts under it: accepted, rejected, and duplicates.

Accepted is the number of samples that made it into the profile. Rejected is the number that did not. Duplicates is the number that repeated something already in the corpus.

Read the three together rather than only the first. A campaign showing three accepted and zero rejected has a clean corpus. A campaign showing three accepted and eight rejected has a problem with what you are feeding it, and adding more of the same will not help. High duplicates usually means you pasted variants of one email rather than distinct emails, which produces a narrower profile than the count suggests.

Under the counts is a Generated with this profile box holding the identifiers for the corpus, the profile, the analysis version, and the instruction revision the profile was generated against. You do not need to read those values day to day, but they are the way to tell whether a profile predates an instruction change. A profile generated against an older instruction revision was built before your current brief existed.

The extracted style itself appears as a row of chips under the boxes, with plain descriptors such as conversational, direct, persuasive, informal, question-oriented, and analytical. Those chips are the fastest sanity check on the card. If they do not describe how you write, the samples were wrong.

Under the counts is a Waiting for outcomes box. It reports how many sends have gone out using the profile and states plainly that their reply window has not matured yet.

This is the honest part of the card. A new profile is an unproven hypothesis about your voice. Sends are being recorded against it, but a reply can arrive days after an email lands, so the system does not claim the profile is working until enough of those windows have closed.

What this means for you is that a profile is not judged on the day you create it. If you add samples on Monday and check the card on Tuesday, it will still be waiting, and that is correct behaviour rather than a stalled job. The number in the box tells you how much evidence is accumulating.

It also means the sensible review point is later. Set the profile up, let the campaign run, and come back when the send count has grown rather than tuning the samples every day. Changing the profile resets what the accumulated sends were measuring.

Outcome evidence at the campaign and workspace level shows up on the Learning page, which reports treatment and control arms and holds back a lift figure until the outcome windows mature. The writing profile follows the same discipline, which is why both surfaces say they are waiting rather than showing a number that would be noise.

At the bottom of the card is a Delete retained samples button with a line above it that states exactly what the button does. Deleting removes the retained source content, and saved instructions remain unchanged.

That separation is the point. Removing samples clears the emails you pasted in and the patterns derived from them. It does not touch the WRITING STYLE field, the campaign strategy, or anything else in the explicit instructions section, because those are a different input.

There are two normal reasons to use it. The first is privacy: a sample contained customer detail you no longer want retained, so you remove the source content. The second is a bad corpus: the chips do not describe your voice, the rejected count is high, and starting over is faster than adding corrective samples on top.

After deleting, the campaign keeps drafting from your explicit instructions alone. Drafts do not stop, they lose the learned style layer. Rebuild by clicking Analyze more and adding a cleaner set.

A Show version history link sits at the bottom of the card. Open it when you need to see how the profile has changed over time, which is the quickest way to tell whether a change in draft quality lines up with a change in the profile or with something else in the brief.