Guide

How to Review AI Generated Work: QA Checklist

By the Space Office team · Updated September 20, 2026 · 8 min read

To review AI generated work, judge it against the brief, facts, audience, format, and risk before you judge whether it sounds impressive. The fastest safe method is a repeatable QA checklist: define the job, verify claims, score completeness, test anything functional, and send precise revision notes. AI can produce quickly; review decides whether the output is usable.

The dangerous thing about AI generated work is not that it is always bad. The dangerous thing is that weak work often looks finished. It has headings, confident sentences, clean formatting, and exactly enough polish to slip past a busy founder. That is why review needs a system, not a vibe check.

Space Office is built around that system. It is a managed team of 24 live AI specialists coordinated by Hydrogen, an AI project manager that reviews every output before delivery. Hydrogen checks outputs against the brief and returns issues to the specialist for revision. You still review and approve the final work.

Start by restating the job in one sentence

The first review step is to restate the assignment so you know what success means. Before reading line by line, write one sentence: “This output is supposed to help [audience] do [job] under [constraint].” If you cannot fill that sentence, you are not ready to review the work.

AI outputs often fail by answering a nearby question. A blog draft becomes a generic essay. A sales email becomes a brand manifesto. A support answer becomes technically correct but emotionally useless. The brief is the ruler; the output is not allowed to define its own target.

Never ask, “Is this good?” first. Ask, “Is this the thing we asked for?”

Use a 100-point scorecard before debating style

A scorecard turns review from opinion into triage. You do not need a complex rubric for every small task, but you do need the same categories every time: fit to brief, factual accuracy, completeness, audience fit, usefulness, format, and risk. A beautiful draft that scores 62 should not ship.

For low-risk internal notes, 80 can be enough. For public pages, customer emails, legal-adjacent content, finance content, or code, raise the bar. The score is not bureaucracy. It is a way to stop the loudest sentence in the draft from winning the review.

Check facts before you edit the prose

1. Pull every claim into the open

Fact-checking comes before copyediting because polished wrong answers are worse than rough correct ones. Pull out every number, date, customer claim, quote, product feature, integration, price, and legal-sounding statement. Then mark each as verified, internal-only, unsupported, or remove.

2. Keep narrow claims narrow

This is where many AI drafts collapse. They may invent a benchmark, modernize an old price, or turn a roadmap item into a live feature. Hydrogen checks outputs against the brief and returns issues to the specialist for revision. You still review and approve the final work.

Review completeness with the reader’s next question

Completeness means the output answers the next practical question a reader would ask. If a guide says “delegate better,” it should show how to write the brief. If a comparison says one tool is better, it should explain for whom, why, and where the other tool still wins.

1. 1List the promises made in the title, brief, or opening paragraph.
2. 2Underline the exact section where each promise is fulfilled.
3. 3Add a note beside every missing promise: answer, remove, or narrow.
4. 4Ask what a skeptical reader would need before acting.
5. 5Send revision notes tied to those gaps, not vague comments like “make it stronger.”

Test functional work like it can break

Functional AI output needs testing, not just reading. For code, run the smallest meaningful gate: typecheck, lint, test, build, or direct reproduction. For spreadsheets, recalculate formulas. For email automations, check the trigger and the unsubscribe path. For website copy, click every link and inspect the mobile layout.

The review question changes from “does this sound right?” to “what would happen if a real user touched it?” That shift catches a different class of mistake: broken links, missing states, wrong assumptions, bad formatting, and outputs that only work in the happy path.

A worked example: scoring a landing page draft

Say an AI drafts a 1,200-word landing page for a SaaS analytics product. The brief asks for founder-friendly language, three use cases, pricing-safe copy, and a waitlist CTA. The draft looks clean, but the review score says otherwise: brief fit 16/20, factual accuracy 12/20, completeness 10/15, audience fit 8/10, usefulness 11/15, format 8/10, risk 8/10.

The total is 16 + 12 + 10 + 8 + 11 + 8 + 8 = 73 out of 100. That is not a publish score. The draft lost 27 points, mostly because it claimed “SOC 2 ready” without proof, skipped one required use case, and used vague value lines instead of concrete examples. The revision note is specific: remove the unsupported compliance claim, add the missing use case, and replace three broad claims with product-specific proof.

Write revision notes an AI can actually use

1. Name the exact failure

Good revision notes are specific, bounded, and testable. “Make this better” is not a review. “Rewrite section 3 for solo founders, remove the unsupported benchmark, add one example with numbers, and keep the section under 180 words” is a review instruction the next pass can follow.

2. Preserve what already works

The best notes name the defect, the reason, and the desired fix. They do not rewrite the whole piece unless the draft is fundamentally wrong. Review should preserve the useful work while forcing the risky or weak parts back through the system.

A useful AI review note says what failed, why it matters, and what must change before delivery.

Set different gates for different risk levels

Not every AI output needs the same review depth. A private brainstorming list can move fast. A pricing page, customer-facing support answer, legal-adjacent article, or code change needs a stricter gate. The mistake is using the same casual skim for everything because the draft arrived quickly.

Where a managed AI team changes the review burden

A single assistant can draft work, but you usually become the reviewer, project manager, and quality gate. Space Office changes that operating model. Hydrogen splits the brief, assigns specialists, and reviews output before delivery, so you make direction calls instead of inspecting every raw draft from scratch.

That does not mean you should stop thinking. It means your review starts later in the chain. You inspect a managed deliverable rather than a pile of disconnected AI attempts. For a founder, that difference matters because attention is the scarce resource.

The practical review checklist

• Restate the job in one sentence before reading closely.
• Score the work against brief fit, facts, completeness, audience, usefulness, format, and risk.
• Verify every number, name, price, customer claim, feature claim, and link.
• Test anything functional instead of trusting the prose around it.
• Send revision notes that say what failed, why it matters, and what must change.
• Publish only when the output is accurate, complete, on-brand, and safe for the audience.

Bottom line: review is where AI becomes work

AI generated work is raw material until it has passed review. The review step catches the nearby answer, the invented number, the missing edge case, the overconfident sentence, and the broken link. If you build one habit, build this: never let speed substitute for a quality gate. Fast drafts are useful. Reviewed work is what you can actually ship.