This is Week 24 of the AI Ops Playbook — the eighth and final week of Round 3 — and it is the trending-AI week, which this time means something a little different from the usual survey of new capabilities. There are now 140 templates in this library. Almost every one of them is a prompt: a thing you paste into a fresh chat to get a specific job done. That format has been the right one, and it will keep being the right one, but it has a property worth naming out loud. A prompt is a fresh start every single time. The model that helps you write a proposal on Thursday has no idea it helped you write one on Tuesday, does not know what you charge, has never seen how you sign off an email, and will cheerfully invent a delivery date if you let it. For a while that was simply the deal. It is not any more. The tools now remember — Custom GPTs, Claude Projects, persistent context — and that quietly changes the unit of work from *a prompt I paste* to *an assistant that already knows*. That sounds like a convenience upgrade. It is actually a category change, and it brings two problems that a prompt never had. Once an assistant holds your business context, other people will use it, which turns a private productivity trick into something closer to a member of staff who has never been told the rules. And once AI is embedded in how the work happens rather than being an experiment you are running, nobody is measuring it any more — because you measure a pilot, and you stop measuring the moment something becomes just how things are done. Persistence, permission, proof. That is this week, and it is the right place to end a round called Run the Business. ## What's New This Week **Build-Your-Own Custom Assistant** (Internal Ops, Advanced) — This builds the thing that should sit underneath everything else in this library: a reusable assistant, a Custom GPT or a Claude Project, permanently loaded with a written brief of your business. The template runs in two phases, and the first one is the real work. It interviews you — offers and pricing, who buys and what they fear, how you actually sound, the claims you cannot make, and the specific facts where an invented answer would cause a real problem — and it pushes back when your answers are vague. Tell it your tone is professional but friendly and it will point out that this describes essentially every business on earth and ask you for two real sentences instead. The output is a context brief of under 1,200 words, and that document, not the assistant, is the asset: it outlives whichever AI tool you happen to be using this year. Phase two turns it into paste-ready instructions, plus the five test prompts that tell you whether the configuration worked. One of those tests matters more than the rest — ask it something the brief does not cover and watch whether it admits the gap or produces a confident invention. **Company AI Usage Policy Generator** (Internal Ops, Beginner) — Your team is already using AI. Not might soon: already, on personal accounts, with whatever they happened to paste in, because nobody has told them where the line is. This writes a plain-language policy for a business with no legal department and no appetite for bureaucracy — approved tools, the data red lines, when a customer has to be told, and who checks output before it leaves the building. It is capped at 900 words on purpose, because a policy nobody finishes reading is a policy nobody follows. Two design choices do most of the work. The permitted list is deliberately longer than the prohibited one, since a document that only forbids things reads as a warning and quietly suppresses the productive use you actually want. And the incident procedure is explicitly blameless, because the realistic risk in a small business is a well-meaning person moving fast, and a policy that punishes reporting simply converts small problems into hidden ones. It also hands you the three decisions you should make deliberately rather than accept as drafted. **AI Workflow ROI Audit** (Strategy & Research, Intermediate) — You have adopted a stack of AI workflows over the past year or two and you have measured precisely none of them. This audits what you are actually running, starting with an inventory phase, because most owners genuinely cannot list their own AI workflows from memory — particularly the ones they set up enthusiastically in March and have not opened since, which are frequently still billing. Then it does the arithmetic properly, counting the two costs that never make it into anyone's mental version: editing time, and verification time. A two-minute draft that takes twenty minutes to fix is a twenty-two-minute task, and it may well be slower than writing the thing yourself. The output is a scoreboard sorted by monthly net value, and then a keep list, a fix list, and a kill list. The prompt is instructed not to soften the kill list. ## Why These Three Together These are not three assorted AI topics. They are the three things that only become problems once AI actually works. Nobody needs a usage policy while they are experimenting, because experiments involve one person being careful. Nobody audits a pilot, because a pilot is already understood to be a bet. And nobody builds a custom assistant for a tool they open twice a month — the effort only pays back at volume, which means the very existence of the question is a signal that the volume arrived. So the honest read of this week is that all three templates are symptoms of success. If none of them feel relevant to you yet, that is useful information rather than a gap: you are earlier in the curve than the templates assume, and the right move is more of the library's ordinary prompts, not governance scaffolding for a thing you are not doing much of. But if you recognized your own business in any of the three descriptions above — the re-explaining, the team using AI on their own accounts, the subscription you are not sure is earning its keep — then you have crossed over, probably some months ago, and nothing about your setup has changed to reflect it. ## The Part That Makes This Hard The uncomfortable one is the audit, and it is worth being direct about why. Most conversations about AI return on investment are, functionally, attempts to justify a decision already made. The spend is happening, the tools are in place, somebody advocated for them, and the analysis exists to confirm that this was wise. That is not an audit. That is a defense. The genuinely valuable output of thirty minutes spent on this is the kill list — and specifically the entry on it that you were quietly proud of. Nearly every business a year into this has two or three workflows where somebody spends twenty minutes editing a two-minute draft into something usable, which is slower than writing it from scratch would have been, and nobody says so. Not through dishonesty. Because AI is supposed to be good, because the workflow was somebody's initiative, and because the generation step feels fast and the editing step feels like normal work rather than like the cost of the generation step. That gap between how it feels and what it costs is the whole reason to write the numbers down. And it is why the template insists you time your editing step once, honestly, rather than estimating it — estimated editing time is optimistic in a completely predictable direction, because you remember the run where the first draft was fine and not the four where you rewrote most of it. Here is the reframe that makes the exercise worth doing rather than depressing. A kill list is not evidence that AI failed you. It is evidence that you tried enough things to have some that did not work, which is the only way anyone ever finds the ones that do. The businesses getting real value are not the ones that picked correctly at the start. They are the ones that noticed what was not working and stopped. ## Round 3 Closes Here Eight weeks, thirty templates, and the library goes from 110 to 140. Round 3 was about the owner's seat, and looking back across it the throughline is clearer than it was at the start: the money you cannot see, the customer feedback you already have and have not read, the content engine, the operating year, AI-native automation, four more industries, the knowledge that walks out the door, and now the infrastructure underneath all of it. Very little of that was about clever prompting. Most of it was about the unglamorous business of making something repeatable. Which is, in the end, the honest summary of what changed in the last two years. The models got better, but the gap between businesses getting real value from AI and businesses generating impressive demos is not a gap in prompting skill. It is context, rules, and measurement. It is the boring part, and this week is three templates for the boring part. ## A Note on Cadence — Thirteen Wednesdays Straight This drop lands on its regular **Wednesday** for the thirteenth consecutive week, an unbroken run from Week 12 through Week 24 and the full length of Round 3. The library is now at 140 templates. ## Get Started All three templates are available now for Pro members: - **Build-Your-Own Custom Assistant** — [open the template →](/templates/build-your-own-custom-assistant) - **Company AI Usage Policy Generator** — [open the template →](/templates/company-ai-usage-policy-generator) - **AI Workflow ROI Audit** — [open the template →](/templates/ai-workflow-roi-audit) Run them in that order if you have an afternoon. Build the assistant first, because the context brief it produces is the input to everything else and it makes every other template in this library work better. Write the policy before anyone else touches what you built. Then put the audit in your calendar for a month from now, once the assistant has had time to be either genuinely useful or quietly ignored, and find out honestly which it was. If you only have twenty minutes today, write the policy. Your team is using AI right now, and the rules you have not written are the ones they cannot follow.