The scenario below is an illustration, not a client account.
A copywriter runs a strong AI session on Tuesday. Good brief, sharp drafts, real edits, something she is happy to send. On the following Tuesday she opens a blank chat and starts again from nothing, having kept none of what the previous week taught her.
Why do good AI sessions not add up to a business that learns?
Because most AI work stops at Produce. The draft appears, it is good enough, and the cycle ends there. Everything that would have made next week start further along, the outcome, the correction, the record of what actually governed the run, never gets captured because the work already felt finished.
The learning happens after Ship. A cycle that ends at Produce generates output. It cannot compound.
The seven stages
Orient → Choose → Specify → Produce → Verify → Ship and measure → Capture and recalibrate
A learning cycle is a workflow that ends with the business knowing something it did not know before, recorded somewhere the next run will read. The draft is the midpoint, not the destination.
Each stage closes a specific gap. That is the useful way to read the list: not as ceremony, but as seven named failures and the step that prevents each one.
| Stage | What it does | The gap it closes |
|---|---|---|
| 1. Orient | Load the smallest relevant slice of business truth. The archive stays out of the run | Work that starts from nothing, or from everything |
| 2. Choose | Name the constraint and the decision the result will inform | Activity impersonating strategy |
| 3. Specify | Turn intent into a brief with criteria that exist before the candidate does | The standard drifting toward the generated answer |
| 4. Produce | Delegate bounded, reversible work | Unbounded delegation you cannot undo |
| 5. Verify | Match the check to the consequence | Fluency becoming trust |
| 6. Ship and measure | Put the smallest responsible version in front of reality | Internal approval becoming market evidence |
| 7. Capture and recalibrate | Route each result to the store whose authority it changes | Paying the same tuition twice |
Stage 2 is where most weeks go wrong
Choose asks three questions, and they are harder than they look.
What belief about the business are we trying to change or test? What decision will the result inform? What is the smallest customer-facing evidence worth obtaining?
If you cannot answer those, production can wait. A month of daily posts and no new conversations is the classic shape of a cycle that skipped Stage 2: high activity, flat learning, and no way to tell which of the two problems you have.
Stage 3 has an ordering rule worth memorising
The brief names the outcome and audience, the authoritative inputs and their priority, the required and forbidden elements, the evidence standard, the output form, the known failure modes, and the fallback path.
The load-bearing part is when. Criteria are written before the candidate exists.
Write them after and they drift, because you have now seen the answer and the standard quietly reshapes itself around what you are looking at. This is not a character flaw. It is what happens to anyone who evaluates work they have already read.
Stage 7 fails, and not because you lack discipline
This is the stage that separates a cycle from a habit, and it is the one that breaks.
It breaks structurally. Stage 7 is the only part of the loop whose execution depends on you choosing to do it after the work is already finished. A step that must be remembered is the first step to disappear under load.
And the runs it skips are not random. Capture degrades exactly on the weeks that were busy enough to produce corrections worth keeping, which biases your learning history toward the days that went well. The record ends up describing a calmer business than the one you actually run.
The architectural answer is not more discipline. It is to make the run produce its own record. Executing a governed workflow should write the ledger, the frozen criteria, the verdict, any override, and the outcome as a by-product of executing, the way a delivery leaves an invoice behind.
What remains yours is disposition rather than transcription: confirming, correcting, redirecting, or rejecting what the run already proposed.
That line matters. Automatic capture may remove your authorship of the record. It must not remove your authority over it. Nothing enters a durable store on the model’s say-so.
Where each result goes
Capture is routing, not filing. Each result goes to the store whose authority it changes.
- A sourced fact or accountable decision about the customer, offer, or channel goes to business knowledge.
- Something the workflow should do differently, in a named scope, for a stated reason, becomes an operating correction.
- A repeatable pattern with evidence beyond one lucky output becomes expertise.
- What happened, what failed, what was overridden, what shipped, and what followed goes to the run ledger.
Putting all four in one file is how a system ends up informed and ungoverned.
The boundary
This is an evidence-informed operating model, not a scientifically proven recipe. Individual mechanisms in the cycle have research support. The cycle as a package is a design proposal.
And one rule survives every stage: a passing verdict does not grant permission to publish, send, spend, delete, or change terms. You approve the consequential action immediately before it happens.
Try this now (5 minutes)
Take the last AI task you finished.
Walk the seven stages and mark each one: ran it, skipped it, or did not know it existed.
Almost everyone marks 1 through 4 as ran and 6 through 7 as skipped. If that is your pattern, you do not have a prompting problem. You have a cycle that ends at the draft.
Then answer one question about the run: what does your business know now that it did not know before, and where is that written down?
If the honest answer is “nowhere,” that is Stage 7, and it is the highest-value stage to fix first.
Where to go next
The authority boundaries the cycle routes into are in the model is not the operating system.
For the smallest possible version of the whole loop, see the smallest governed AI workflow.
For Stage 5 specifically, how to measure AI output quality covers checking against a standard you set.
Frequently asked questions
Do I have to run all seven stages every time? No. Match the depth to the consequence. A private brainstorm does not need a shipped test and a measured outcome. Anything that reaches a customer, changes a price, or makes a claim should run the whole cycle, because those are the runs where a missed stage costs something. The discipline is knowing which stage you skipped and why, rather than skipping by default.
Which stage do people actually skip? Stage 7, Capture, and it fails for a structural reason rather than a motivational one. It is the only stage whose execution depends on choosing to do it after the work is already finished. A step that must be remembered is the first step to go under load, and the weeks it gets skipped are the busy ones, which are exactly the weeks that produced the corrections worth keeping.
If capture keeps failing, should I automate it? Automate the recording, never the authority. A governed run should write its own ledger, frozen criteria, verdict, overrides, and outcome as a by-product of executing, the way a delivery leaves an invoice. What stays with you is disposition: confirming, correcting, redirecting, or rejecting what the run proposed. Automatic capture may remove your authorship of the record. It must not remove your authority over it.
Is this a proven productivity method? No. It is an evidence-informed operating model, and the distinction is deliberate. Individual mechanisms in it have research support. The cycle as a package is a design proposal, not a validated formula, and nothing here predicts a result for your business. Treat it as a structure for finding out, not as a promise about what you will find.