Configura las instrucciones de IA de una campaña

Open the campaign and click the AI Instructions tab in the builder, between Settings and Workflow. The header reads "Custom AI Instructions" with one line under it: everything here is appended to the AI's user prompt for this campaign. A Save now button sits at the right of that header, and the builder shows "All changes saved" at the top when your edits have landed.

Under the header are three numbered steps: 1 Describe campaign, 2 Teach the AI, and 3 Practice your style. They are tabs, not a wizard. You can move between them in any order and nothing forces you through step 3 before launching.

The order still matters for a first pass. Step 1 generates a draft brief from four short answers, step 2 is where you correct that draft and see what the AI will actually be held to, and step 3 lets you sample the result. Skipping straight to step 2 and writing everything by hand works, it is just slower.

Instructions here apply to this campaign only. They do not change other campaigns and they are not the same thing as Chat memory, which holds workspace-level facts. A campaign that needs a different audience or a different offer needs its own instructions.

If the campaign is running, editing instructions is still allowed. The header keeps showing "All changes saved" as you type, and the Pause button stays available if you would rather stop sending while you rework the brief.

Step 1, Describe campaign, opens with a short instruction: start with the four prompts below, they generate a first pass you can edit in step 2 before anything is final. Under it is an Auto-fill card that repeats the important part. You answer four short questions, it fills in editable guidance, English is used by default, and nothing is final until you save the instructions.

Click Auto-fill from a few answers and work through the questions. Answer them the way you would brief a new SDR, in plain sentences with real nouns. "Heads of RevOps at B2B SaaS companies running outbound with two or more reps" produces a usable brief. "Growth-focused companies" produces a brief you will rewrite by hand anyway.

What comes back is a draft, not a decision. Every field it populates is editable in step 2, and the wording it chooses is often close but generic in exactly the places that matter, such as the offer and the call to action. Treat auto-fill as a way to avoid an empty form rather than as the final brief.

Auto-fill is safe to run more than once while you are still setting up. Nothing is committed until you save, so if the first pass came back too broad, answer the questions again with a narrower audience rather than editing a bad draft field by field.

English is the default language for the generated guidance regardless of the language your prospects speak. If you send in another language, plan to rewrite the writing style and subject guidance in step 2.

Click 2 Teach the AI. Scroll past the writing profile and testimonial cards to the Explicit instructions section. Its description sets out the relationship between the two inputs: this is where you refine the guidance generated in step 1 or write your own, and the approved writing profile above stays a separate learned input so neither source silently overwrites the other.

The fields are specific, and each one has a job.

• CAMPAIGN STRATEGY defines who this campaign is for, why the offer is relevant, and the single action each email should work toward.
• TARGET AUDIENCE names the reader, with a placeholder showing the intended shape: revenue leaders at B2B SaaS companies with a growing outbound team.
• EXCLUSIONS and RELEVANT TRIGGERS take one entry at a time, for audiences to keep out and events worth referencing.
• PROBLEMS holds the problems this campaign can address.
• OFFER and BOUNDED DELIVERABLE separate what you are selling from what you can safely promise in a first email.
• DESIRED ACTION and CTA POLICY set the one action you want and how the ask should adapt to the recipient's apparent intent.
• WRITING STYLE covers sentence length, formality, structure, and signature quirks.

Fill BOUNDED DELIVERABLE carefully. It is the field that stops drafts promising something you cannot deliver, and leaving it empty is how a campaign ends up offering an audit you never agreed to run.

The right column of step 2 holds the EFFECTIVE BRIEF pane, with a Saved marker at its top right. This is the compiled result of everything you entered, and its first line states its role: generated drafts are checked against these explicit constraints, and violations block readiness until fixed.

The pane is organised as a document. ### Campaign carries the objective and target audience. ### Voice carries the sender identities, the writing style in full prose, subject guidance, and formatting rules with real numbers, for example sentence length in a stated word range and paragraph length in a stated sentence range. ### Constraints lists the explicit user instructions that are never overridden, split into a Do list and a Don't list.

Read this pane rather than trusting your field entries. It shows what the drafting step is actually held to, including how your free-text writing style was interpreted. If you wrote a style note and it does not appear here, it did not make it into the brief.

The constraint lists are the highest-value part to check. Entries like never mentioning internal tooling, never copying website wording verbatim, and never inventing recipient-specific facts are the rules that keep a draft from embarrassing you. If your industry has a rule of its own, add it in the explicit fields and confirm it appears in this list before launching.

Because violations block readiness, a strict brief has a real cost. Drafts that cannot satisfy it will not go out, which is the intended behaviour but can look like a stalled campaign if you did not expect it.

Under the effective brief is a CONFLICTS RESOLVED panel with a count next to it. It lists every place where two instructions collided and records which one won.

Each line names the loser, the winner, and the reason. A compliance rule overriding an explicit preference reads as a compliance override of the user preference, with the reason attached, for example a protected opener contract that does not allow dash punctuation in greetings. A newer instruction overriding an older one reads the other way, with the current campaign instructions superseding a legacy workflow instruction.

Two patterns are worth reacting to. The first is a compliance override, which tells you a safety or formatting contract beat something you asked for. Your instruction was not ignored, it was outranked, and rewording it will not change the outcome. The second is the same conflict repeating several times, which usually means one phrase in your instructions keeps colliding with a protected rule. Fix that phrase once rather than accepting six identical overrides.

An empty or short conflicts list is the healthy state. A long one means you wrote instructions the system cannot honour, and every one of those is a place where drafts will not read the way you expect.

Check this panel after every meaningful edit to the explicit fields, not just before the first launch. New instructions can create new conflicts with rules that were never in tension before.

Click 3 Practice your style. The panel explains itself in one line: generate a few practice emails written in the style above, then keep the ones that sound like you, and this never blocks launching your campaign. A Generate practice emails button sits under it, and a counter at the top right of the tab tracks how many you have confirmed out of the available set.

Practice emails are the cheapest way to catch a brief problem before real prospects see it. Generate a batch and read them end to end rather than skimming subject lines. The failure mode you are looking for is a strong opener followed by a generic middle, which usually means PROBLEMS or RELEVANT TRIGGERS is too thin.

Keep the drafts that sound like you and let the rest go. Confirming a draft you do not actually like is worse than confirming nothing, because it teaches the wrong target.

The counter is a progress indicator rather than a gate. The step is optional by design and the campaign launches with zero confirmed practice emails, so use it when the campaign matters and skip it when you are testing plumbing.

If practice emails come back consistently wrong in the same way, go back to step 2 rather than generating more. The brief is the input, and more samples of a bad brief only produce more bad samples.