Lifecycle content engine with a brand gate (L3)

The steps

1. 01

   Audit the journeys, not the content calendar

   Tool: Braze

   Open every live automated journey and mark each step as one of three things: missing content entirely, running content older than 12 months, or using one generic message across multiple segments. That marked list is your work queue. The content calendar is not the work queue, the calendar is what someone planned, the journey audit is what is actually broken. Rank by volume of contacts flowing through each step, so the highest-traffic gap gets fixed first. • Owner: Lifecycle marketer • Tool options: the journey builder in Braze, Klaviyo, Iterable, Marketo or HubSpot • Pitfall: treating the content calendar as the source of truth instead of the journey audit • Definition of done: a ranked list of content gaps exists with journey, step and segment named, sorted by contact volume

2. 02

   Build a voice file the model reads every time

   Tool: ChatGPT

   The file contains six sections: ten approved sentences you love pulled from real shipped work, ten sentences you would never publish each with one line on why (these do more work than the good ones), banned words and phrases as an explicit list, claim rules covering what you may say about results, what requires a disclaimer and what legal has already cleared, reading level and length targets per format, and approved CTA formats verbatim. Load it into the Project's instructions so it applies without anyone remembering to paste it. A voice guide nobody has to remember is the only kind that works. Give it an owner and a quarterly review date. • Owner: Brand or content lead • Tool options: one markdown file loaded into a ChatGPT Project, Claude Project or Gemini Gem • Pitfall: relying on writers to remember to paste the voice file instead of loading it into the Project instructions • Definition of done: two different writers using the Project produce outputs a brand lead cannot tell apart in a blind test

3. 03

   Generate against a skeleton, never a blank request

   Tool: ChatGPT

   Every request gives the model the outline of the output rather than a vague ask. Six inputs, every time: the segment, the trigger that put them here, the one job this message does, the single proof point available, the CTA, and the length. Give it the skeleton and it fills in the muscles. A blank "write me a welcome email" produces average content, because average is what a model returns when you have not told it what specifically to be. • Owner: Lifecycle marketer • Tool options: the Project plus the journey audit • Pitfall: sending vague, blank requests instead of the six-input skeleton • Definition of done: 10 pieces are drafted across the top three gaps, each traceable to a specific journey step

4. 04

   Route through the identical approval path, with no fast lane

   Tool: Manual

   AI-drafted content goes through the same reviewers and the same SLA as human-written content. No expedited path because it was cheap to produce. Tag it AI-assisted in your internal metadata for audit purposes, and never externally. If review becomes the bottleneck, that is real information, it means step 6's cap is set too high, not that review should be relaxed. • Owner: Content lead plus legal • Tool options: whatever review process you use today • Pitfall: giving AI-drafted content an expedited path because it was cheap to produce • Definition of done: 10 pieces have cleared review with internal tagging in place

5. 05

   Ship every piece as an A/B test against the incumbent

   Tool: Looker

   New AI-assisted content always launches against the current message, never as a straight replacement. Declare the success metric and the minimum run length before launch, two weeks minimum, no peeking-and-stopping when it looks good on day three. When a variant loses, archive it with one written line on why you think it lost, and paste that line into the voice file. The losses are how the voice file gets good. • Owner: Lifecycle marketer • Tool options: the experiment feature in Braze, Klaviyo, Iterable, Marketo or HubSpot; results read in Looker, Tableau or Power BI • Pitfall: shipping AI content as a straight replacement instead of a declared A/B test • Definition of done: three tests have completed with declared winners and the voice file has been updated from at least one loss

6. 06

   Set a monthly production ceiling equal to review capacity

   Tool: Manual

   Calculate how many pieces review can genuinely absorb per month, then cap net-new production at that number. Write the cap down and hold it. Unbounded generation produces a review backlog, then pressure to wave things through, then unreviewed content in front of customers. The cap is a feature, and the tell that you need it is output climbing while engagement falls. • Owner: VP Marketing • Tool options: the content calendar • Pitfall: letting generation volume climb past what review can genuinely absorb • Definition of done: a written monthly cap exists and has held for two consecutive months

7. 07

   Instrument the result

   Tool: Looker

   Instrument: conversion lift of AI-assisted content versus incumbent, plus cycle time from brief to live. The second number is why anyone agreed to this; the first is whether it worked. • Where it breaks: volume goes up five-fold, review capacity stays flat, and generic content ships because someone waved it through under deadline. Second failure: no voice file, so every writer's output sounds like a different company. Third: content ships as a replacement instead of a test, so you never learn whether it was better than what you already had. • Visual guidance: a content pipeline with a deliberate narrowing at the review stage, labeled with the actual monthly cap number. The bottleneck is the design, so draw it as one.

Tools in this playbook

• Braze
• ChatGPT
• Manual
• Looker

Next playbooks

• AI customer-interview synthesis (L3)
• AI SEO clusters with Clearscope/Frase (L3)
• AI-generated event follow-up (L3)
• Lifecycle email AI personalization (L3)