Quality control so a bad output never reaches a client

The single feature that separates a fleet you can trust from a fleet that will eventually humiliate you is a reviewer agent that reads every output against a fixed standard before a human ever sees it, because speed without a quality gate just ships your mistakes faster.

Your fleet now produces work at a pace you could never match by hand. That is exactly the danger. An agent that drafts a wrong figure, invents a fact, or misreads a brief will do it confidently and fast, and without a gate it will reach a client before you notice. Quality control is the role that turns a fast fleet into a safe one, and it is the role most people skip until a bad output costs them a client.

The reviewer agent and its fixed standard

A reviewer agent checks the drafter's output against a written checklist, the same checklist every time, and either passes it or sends it back with the specific failures named.

The power is in the fixed standard. A human reviewing their own work is inconsistent, kind to themselves when tired, and blind to the error they just made. A reviewer agent runs the identical checklist on every output: does every figure trace to the source brief, is every claim supported, does the tone match the client's, are there forbidden moves (a promise we cannot keep, a number we cannot verify). Because the checklist is explicit and external, the standard does not drift with your mood. Write the checklist once, as the reviewer's definition, and improve it every time a mistake slips through.

The reviewer never writes and never sends. Its only job is to judge against the standard and route accordingly: pass to the human gate, or return to the drafter with named fixes. That separation is what keeps it honest, because a reviewer that could also rewrite would start excusing the work it is meant to catch.

The "would a client notice if this were wrong" test

The checklist's top question is the one that protects your reputation: would a client notice, and care, if this specific thing were wrong?

Sort every check by that test. A wrong headline figure in a performance report: a client would absolutely notice, so it is a hard-fail check. A slightly clumsy sentence: a client probably would not care, so it is a soft note, not a blocker. This sorting stops the reviewer from blocking on trivia while waving through the errors that actually cost you. The reviewer's hard-fail list should be short and brutal: the handful of mistakes that would make a client doubt your competence. Everything else is advisory.

The reframe worth holding: quality control is not about making every output perfect, it is about guaranteeing that no output with a reputation-damaging error reaches a human. Perfect is the enemy of shipped; safe is the actual target.

The spot-check cadence and what you never automate fully

A reviewer agent is not infallible, so you sample its work on a fixed cadence rather than trusting it blindly.

Pick a rate (one in ten passed outputs, or every output for a new client) and read those yourself against the same checklist. When you find something the reviewer missed, you do not just fix that output; you add the miss to the reviewer's checklist, so the whole class of error is caught from then on. That is how the gate gets stronger over time instead of decaying. And there are outputs you never let any agent send unreviewed regardless of how much you trust the chain: anything that goes to a client, anything that moves money, anything legal or contractual, anything that cannot be unsent. Those always pass through both the reviewer agent and your own eyes.

A worked example: the gate that caught the errors before send

Take a freelance financial copywriter producing investor-update emails for five startup clients, where a single wrong number could undermine a founder's credibility with their investors.

• Challenge. The drafter agent wrote fluent updates fast, but across a month it produced several updates with figures that did not match the source data: a growth rate transposed, a revenue figure from the wrong month. Any one of them reaching investors would have been a serious problem for the client.
• Approach. She added a reviewer agent with a hard-fail checklist: every number in the email must trace to the source brief, every percentage must be recalculated from the underlying figures, and any claim about performance must be verifiable. The reviewer returned failing drafts to the drafter with the exact mismatched figure named, and only clean drafts reached her own final human gate.
• Result. Over the next quarter the reviewer caught 11 figure errors before any reached the human gate, and zero wrong numbers went out to investors across roughly 60 updates. The drafter's raw error rate did not change; the gate simply caught every one before send, which is the only outcome that matters to the client.

Pitfalls in quality control

• Letting the reviewer also rewrite. A reviewer that fixes its own findings stops being a check and starts excusing the work. Keep judging and writing in separate agents.
• A checklist that blocks on trivia. If the reviewer fails drafts for clumsy phrasing, you will start ignoring it. Reserve hard-fails for errors a client would notice and care about.
• Trusting the gate without sampling. A reviewer agent has blind spots. Spot-check a fixed share of passed outputs and feed every miss back into the checklist.
• Automating the un-unsendable. Client sends, money movements, and legal text always get both agent review and your own eyes, no exceptions.

INTERVIEW EWOUD: Tell the story of the QA gate that saved a client deliverable. What did the reviewer agent catch that you would have missed, and what did you add to the checklist afterwards so that class of error could never slip through again?

With a reviewer gate protecting every output, the fleet is safe to run at speed. What remains is the habit that keeps it running well: the daily and weekly cadence of conducting it. The next chapter sets out the operating rhythm of a one-person fleet.