Stop Asking One AI Chat to Do Four Jobs

The fastest way to get a confident, polished, incomplete result is to ask one conversation to research the facts, make the artifact, review its own work, and declare itself finished.

That is not primarily a prompting problem. It is a division-of-labor problem.

I kept treating AI work as one continuous act: explain the task, wait for the answer, request a few changes, and copy the result. It felt efficient because everything stayed in one place. It also made four different failure modes almost impossible to see.

• Research could be stale.
• Production could drift from the requested artifact.
• Review could defend the maker's assumptions.
• "Done" could mean nothing more than a fluent paragraph saying the work was complete.

The fix was not a longer master prompt. It was four visible passes.

Pass 1: Scout

The Scout does not make the final artifact. Its job is to find the ground the artifact must stand on.

For a research task, that means direct sources, current dates, contrary evidence, constraints, and unknowns. For a content task, it means supported facts, real reader questions, and examples that can be used without inventing a story. For a digital product, it means problem evidence, buyer language, existing alternatives, and the gaps you still cannot prove.

The Scout should label the evidence:

• Verified — checked against a current direct source.
• Supplied — provided by the operator but not independently checked.
• Inference — a conclusion drawn from the available facts.
• Unknown — missing and unsafe to guess.

This single step prevents the Maker from turning uncertainty into polished fiction.

Pass 2: Maker

The Maker creates one bounded artifact.

Not "help with competitor research." A one-page comparison of five current alternatives, with every price cited and one positioning decision.

Not "write launch content." One landing-page outline and three factual posts using only the verified product scope.

Not "improve this process." A checklist with a trigger, input, output, owner, exception path, and completion receipt.

The Maker needs an explicit finish line:

FINISHED ARTIFACT
The exact file, page, message, decision, or delivered item that must exist.

ACCEPTANCE CHECKS
The observable tests that make it complete.

BOUNDARIES
What must not be changed, disclosed, invented, or executed.

Without these three fields, more generation usually means more material, not more completion.

Pass 3: Skeptic

The Skeptic is not an editor asked to "make it better." That instruction often produces smoother prose while preserving the underlying mistake.

The Skeptic tries to disprove completion.

It asks:

• Which claim is stronger than its evidence?
• Which required part is missing?
• Which link, price, date, filename, or total is inconsistent?
• Which example could be mistaken for a real customer or result?
• Which private detail is unnecessary for the public artifact?
• Does the call to action open the page it describes?
• Is there evidence that an external action actually happened?

For public, customer-facing, reputational, financial, or privacy-sensitive work, this should be a separate pass with the original acceptance criteria in view.

The most useful review request I have found is short:

Do not rewrite yet.
List only defects that prevent this artifact from being accurate,
usable, safe, or complete. Tie each defect to an acceptance check
and propose the smallest verifiable fix.

Pass 4: Finisher

The Finisher applies the accepted fixes, packages the result, and records proof.

This is the pass most AI workflows quietly skip. A draft exists, so the work is described as shipped. A deployment command succeeded, so the public page is assumed to work. A post was prepared, so it is counted as published.

The Finisher requires a receipt appropriate to the task:

• a canonical file that opens;
• a tested archive whose contents match the description;
• a public URL checked outside the editor;
• a delivery receipt with known limitations;
• or an explicit failed/partial result when completion was not achieved.

The rule is simple: the statement "done" is never the evidence that the work is done.

The Handoff Between Passes

The four-pass system does not require four different products or four permanent agents. One assistant can perform several roles. The important part is that the passes and their responsibilities stay separate.

The handoff should be smaller than the conversation:

OBJECTIVE
CURRENT VERIFIED STATE
SOURCE OF TRUTH
DECISIONS ALREADY MADE
OPEN WORK
ACCEPTANCE CHECKS
DO NOT

Pass only details that change the next role's decision, output, or safety boundary. Personal backstory, unrelated account history, stale attempts, credentials, and private messages do not belong in a routine handoff.

When One Pass Is Enough

Not every task needs ceremony.

A reversible internal draft using supplied facts can stay in one session. Keep the stages separate, but the same context can Scout, Make, check itself once, and finish.

Split the roles when:

• facts may have changed;
• the output will be public or customer-facing;
• a missed error affects money, reputation, or privacy;
• the task includes an external or irreversible action;
• or "looks good" is not a sufficient acceptance test.

The goal is not maximum process. It is the smallest process that makes the expensive failure visible before it escapes.

A Copyable Role Card

SCOUT MUST FIND:

MAKER MUST CREATE:

SKEPTIC MUST TRY TO DISPROVE:

FINISHER MUST VERIFY:

HUMAN-ONLY DECISIONS:

STOP CONDITIONS:

Write the finished artifact first. Then fill the six lines. If the roles still sound vague, the task is not bounded enough to route.

I turned this card into a free browser-only AI Team Role Router. It generates the four passes locally, without signup or uploading the answers.

The complete $19 AI Solo Operator System adds outcome scoping, handoff packets, quality gates, privacy and proof rules, six real-work playbooks, and a full product-validation module. The templates provide structure; they do not guarantee traffic, customers, or error-free output.

THE OPERATING LOOP

Scout the evidence. Make one bounded artifact. Ask a Skeptic to try to disprove completion. Let a Finisher apply the fixes and record the receipt. Fluency is not proof, and one long conversation is not a division of labor.