Skip to content

AI Systems

Controlled generation pipelines: catch it at the stage that made it

One prompt that writes a whole page gives you a draft you can only accept or redo. A pipeline breaks the work into stages: brief, sources, outline, draft, checks and review. Each has one job, saves its output and is checked before the next begins, so an unsupported claim is caught and sent back to the stage that wrote it. The result is drafts an editor can approve instead of rewrite.
Discuss your systemFollow one page through ↓
BriefC-114Sources4 approvedOutline5 sectionsDraftv1 → v2Checks5 rulesEditorreviewno source · back to draftDraft · Water heater replacement · Eastsidev1v2We replace gas and electric tank water heaters across Eastside.S1Same-day installation, guaranteed.Same-day installation, guaranteed.no sourceInstalls are usually booked within two business days of approval.S2Old units are taken away and recycled.S3checksevery claim has a sourceservice names from the liststyle rulesnot a near-copyheadings match the outlineapproved · editor

One prompt, one shot

Ask a model for a whole service page and you get something fluent. Somewhere in it may be a promise you don't make, a service you don't offer or a town you don't cover, and you can only find it by reading every line. When you fix it and ask again, something else changes.

A pipeline makes each step small enough to check. The model never decides what's true: it works from sources you approved, and names, prices and addresses come from your records. When something is wrong, you can see which stage made it and rerun only that stage.

Follow one page through the pipeline

The Eastside water heater replacement page, stage by stage. Each stage saves what it produced.

Stage 1 of 6: Brief

brief C-114

page
Water heater replacement, Eastside
reader
a homeowner whose water heater is failing
must answer
what we install, how soon, what affects the cost, what happens to the old unit
sources allowed
S1–S4 only
length
450–650 words
never
prices, or promises the policy doesn't make

What's checked between stages

Rules in code run first. A model is used only where the check is about meaning.

  1. Every claim has a source

    Supported, or removed

    Each claim points at an approved source, and the source has to say it.

  2. Names from your lists

    Services, places, products

    Service names and locations must match your lists exactly, so nothing you don't offer appears.

  3. Style and banned claims

    Your rules, written down

    Tone, reading level and phrases you never use, like guarantees the policy doesn't make.

  4. Not a near-copy

    Each page earns its place

    The draft is compared with existing pages so ten locations don't get the same page with a different name.

  5. Structure matches the brief

    Headings, length, fields

    The output has the sections, length and fields the brief asked for, or it goes back.

The model writes. Code handles the facts.

Outlinea model, from the briefOrdering ideas is language work
Paragraphsa model, from approved sources onlyFluent text, from facts you've approved
Names, prices, addressescode, from your recordsFacts are copied, never written
Links and formattingcodeThe same every time
Checkscode first, a model where meaning mattersRules are cheaper and more predictable
Approvala personResponsibility stays with you

What a pipeline hands off

In

Brief
reader, purpose, limits
Approved sources
the only facts allowed
Rules
style, terms, banned claims
Stages
outline → draft → check → revise

Out

Draft
every claim with its source
Check record
what passed, what went back
Review task
for an editor

Worth building when you make many pieces of the same kind

Good fit

  • You produce many similar pieces: service pages, product descriptions, client reports.
  • The facts have to come from your records, not from the model.
  • Someone will own the rules and review what comes out.

Not the right fit

  • One-off writing. A person working with an AI assistant is faster.
  • There are no approved sources to work from. The pipeline would just generate confidently.

Where pipelines do the writing

Next to controlled generation

Questions about controlled generation

Isn't this just a longer prompt?

No. Each stage is a separate call with its own instructions, input and check, and its output is saved. When something is wrong you can see which stage made it and rerun only that stage.

Can it publish without a person?

For low-risk content, once the checks have proven themselves on your own cases. We start with a person approving everything.

How do you stop it inventing facts?

The draft stage sees only approved sources, every claim has to point at one, and a check rejects claims that don't. Names, prices and addresses are inserted by code from your records.

Does every stage use the same model?

No. Each stage uses whatever does that job well. A smaller model can often run the checks.