The Secret to Perfect AI Outputs: Name the Canon
Nearly every established discipline has a canonical reference that already settled the argument you're having with a model. Name it, hand it over, and let the model regularize its own work against a standard instead of your improvised feedback.
By Geordie Everitt
Every model has an opinion about how your work should look, and that opinion is an average, the statistical center of mass of everything it read. An average is nobody's actual standard. It's the compromise position that emerges when no particular authority was named, which is most of the time, because most people asking a model to write, draft, or format something never think to name one.
Most established crafts didn't leave this to chance. Journalism has the AP Stylebook. Academic writing has Chicago, or MLA, or APA, depending on the field, each one a genuinely opinionated document that took a position and defended it across dozens of editions. Project management has the PMBOK Guide, which defines exactly what has to be in a charter, a risk register, or a work breakdown structure before a document earns the name. Law has the Bluebook, four hundred pages settling exactly how a citation gets punctuated so two lawyers in different states produce interchangeable references. Most professional certifications publish their own version of this: an exam content outline or a body of knowledge, written for the specific purpose of getting two people who never met to converge on the same material.
These documents exist because the argument about the right way to do the thing already happened, sometimes across a century, and the canon is what survived it. That's not a small saving. It's the compressed output of professional disagreement that a field already fought and settled, available for free, sitting on a shelf, in most cases entirely public domain or close to it.
What the model does with an opinion it's actually given
Ask a model to fix your prose and it fixes it toward its own average. Ask it to fix your prose according to Strunk and White, or the house style at The Economist, or the Chicago Manual's stance on the serial comma, and it stops averaging and starts applying a rule someone actually defended in print. The difference isn't subtle. An averaged edit second-guesses itself paragraph to paragraph, because there was never a fixed target to converge on. An edit against a named canon gets consistent, because the target didn't move.
This works because the canon is itself heavily represented in what the model was trained on. A model that has ingested the AP Stylebook, or thousands of documents written to comply with it, already contains the pattern. What's missing is usually not the knowledge. It's the instruction to apply that specific knowledge instead of the ambient average sitting on top of it. Naming the canon is how you reach past the average and pull out the thing actually trained underneath it.
Where this gets underused
The obvious cases are the writing ones, style guides, citation formats, the register question a Try This post already covered. The underused cases are the ones more people actually run into during an ordinary week. A project status report built without the PMBOK's own definitions of scope, schedule, and risk baked into the request is prose that happens to borrow PM vocabulary, not a document a sponsor can act on. Study material for a certification exam built without the current exam content outline named is optimizing for a plausible-sounding test instead of the actual one. A compliance document, a HIPAA policy, a SOC 2 control narrative, an ISO 27001 statement of applicability, built without naming the actual framework is a draft an auditor will read once and hand back.
The financial case needs one more distinction, because it's easy to overreach with it. GAAP and IFRS are the wrong tools for generating a balance sheet's actual figures. Those numbers belong to a real ledger, not a model's plausible reconstruction of one, for the same reason a spreadsheet total shouldn't come from a model guessing at arithmetic instead of computing it. What the standard is legitimately good for is the second pass: checking that a number a deterministic system already produced got classified, disclosed, and organized the way the canon requires. Use it to audit the output of a real system of record. Don't use it to invent the output in the first place.
In every one of these cases, the fix costs one sentence: name the standard before you ask for the work, not after you've gotten something wrong back and are trying to explain, from scratch, in your own improvised words, what the standard would have said anyway. You are not smarter than the canon at explaining the canon. Let the actual document do that job.
The move
Before asking for a piece of work in an established craft, spend thirty seconds asking a different question first: what's the canonical reference here, and has this field already settled the argument I'm about to have with the model. Usually it has. Name the document, hand it the authority, and let the model regularize its own output against a standard instead of against your mood that day.
The trick is the oldest one in every one of these fields. What's new is having something in the loop fast enough, and well-read enough, to actually apply it on the first pass.