Quality Control & Review

Not everything needs the same scrutiny, and pretending it does either wastes hours or, worse, trains people to skim everything equally. The organising principle of review is the one that runs through the whole course: match your scrutiny to the stakes. A discarded brainstorm and a specification clause that goes to site are not checked the same way, and a good review discipline says so explicitly.

Sort Claude-assisted output into rough tiers. Low-stakes, divergent work - mood-words, rough option ideas, an internal summary - needs a sanity glance: is it useful, roughly right, on-brief? You were going to curate it anyway. Medium-stakes work - a client email, an internal report, a first-draft proposal - needs a proper read for accuracy, tone and completeness before it goes out, because it carries the practice's name even if it is not legally binding. High-stakes, convergent work - anything with a number, a dimension, a clause, a standard, a legal or safety implication, anything that goes to site or into a contract - needs line-by-line verification against the actual source, by someone competent to judge it. Here a confident-wrong answer is a real liability, and Claude's fluency makes it more dangerous, not less.

The practical move is to make these tiers visible in how the studio works. When a document is Claude-assisted, note what tier it is and what was checked - a light convention, not bureaucracy. It stops the specific failure that sinks practices: high-stakes output riding out on the confidence earned by low-stakes output, because "Claude's been great all week." Consistency of quality does not come from trusting Claude uniformly; it comes from distrusting it in proportion to what a mistake would cost. That single habit, taught in the rollout (9.3) and enforced at review, is the difference between a practice that is faster and one that is merely more exposed.

A checklist turns "read it carefully" from a good intention into a repeatable act - and, like the surgical checklist it borrows from, its value is that it is boring and does not skip the obvious under pressure. Build one for your studio; here is a spine to adapt, worked hardest on the high-stakes tier.

BEFORE ANY CLAUDE-ASSISTED WORK SHIPS
[ ] Every figure, dimension, quantity checked against source or by hand
[ ] Every standard / code / clause number confirmed to exist AND to say
    what the text claims
[ ] Every cited fact, product, precedent or reference verified real
[ ] Names, dates, project specifics correct (Claude drifts on these)
[ ] Nothing invented to fill a gap - [VERIFY] flags all resolved
[ ] Reads in the practice voice; on-brief; complete (nothing omitted)
[ ] A competent human has read the whole thing, not skimmed
[ ] Confidential data handled correctly for the plan used

The first five lines target Claude's specific weaknesses; the rest catch the ordinary ways any draft fails. Two points about using it well. First, verification means going to the source, not asking Claude "are you sure?" - it will often cheerfully confirm a fabrication, because agreeing is also plausible. A confirmed standard is one you found in the standard, not one Claude restated. Second, the human read is non-negotiable and non-delegable to the tool: the checklist ends with a person, because a person is who signs.

Adapt the weight by tier. Low-stakes work might touch only the last three lines with a glance. High-stakes work runs the whole list, line by line, by someone qualified in that domain - the technical lead for a spec, whoever owns the numbers for a fee letter. Keep the checklist in the shared system beside the Projects and prompt library, so it is part of how the studio works rather than a poster nobody reads. And treat a caught error as a gift to the library: if Claude keeps making the same mistake, add a line to the prompt that pre-empts it and a line to the checklist that hunts it.

So far Claude has been the thing under review. But it is also an excellent reviewer of human work - a fast, tireless, well-read second pair of eyes that never gets bored on page forty of a report. This is one of its most valuable and underused roles in a studio, and it has a crucial safety property the reverse does not: when Claude checks your work, you remain the expert who judges its suggestions, so its fallibility is contained.

Use it to pressure-test what you have written. Paste your draft specification and ask it to find internal contradictions, missing sections, ambiguous clauses and undefined terms. Give it your fee proposal and ask what a wary client would query. Hand it your report and ask for gaps in the argument, or your drawing notes and ask what a contractor might misread. It is genuinely good at this - a fresh, comprehensive read against a brief you supply - and it catches the ordinary human errors of omission and inconsistency that a tired author misses.

Here is our draft finishes specification [paste]. Acting as a careful
reviewer, list: internal contradictions; anything ambiguous or open to
misreading on site; sections a complete spec should have but this lacks;
and any term used but not defined. Do NOT rewrite it - just flag, with
location, so I can judge each one.

Notice the framing: it flags, you judge. That is the whole discipline. Claude as reviewer surfaces candidates for your attention; it does not adjudicate, because its "this is fine" carries no more truth-guarantee than its "this is wrong." Two honest cautions. It will sometimes miss a real problem (a false all-clear) and sometimes flag a non-problem (a false alarm), so a clean review from Claude is reassurance, never a certificate. And a second read by the same kind of engine has correlated blind spots - Claude reviewing Claude is not two independent eyes. For high-stakes work, Claude's review supplements a human reviewer; it never replaces one. Used within those limits, though, it measurably raises the quality of what leaves the studio, and it does so on work no human colleague had the hours to check twice.

You review better when you know what you are hunting. Claude's mistakes are not random; they cluster in predictable places, and a studio that names them catches far more than one that reviews vaguely. Teach this list in the rollout and keep it beside the checklist.

Fabricated specifics. Invented standard numbers, clause references, product codes, citations and statistics - stated with the same fluency as true ones. This is the classic hallucination, and it is most dangerous exactly where it matters most: codes, specs, legal and technical text. Every specific gets verified at source.

Wrong maths. Claude can set up a calculation well and still get the arithmetic wrong, especially across multiple steps or unit conversions. Treat any number it produces as a draft to check by hand or a real tool - never a quantity survey, never a signed-off figure.

Plausible-but-wrong substance. The subtlest failure: a spec clause that reads perfectly but is not quite right for your assembly, a summary that quietly misstates the source, advice that is generically sensible but wrong for this site or code. Fluency hides it, which is why domain expertise, not proofreading, is what catches it.

Drift on specifics. Names, dates, project numbers, which room is which - Claude can transpose or invent these while getting the surrounding prose right. Cross-check every project-specific detail.

Outdated knowledge. Without web access it relies on training with a cutoff and will not know the newest code amendment, product or event - and may confidently give superseded information. Confirm anything time-sensitive against a current source. (The withdrawal of an old code edition in favour of a new one is exactly the kind of thing it can miss.)

Sycophantic agreement. Push back and it often folds and agrees, whether or not you were right; ask "are you sure?" and it may reverse a correct answer or confirm a wrong one. Its agreement is not evidence. Verify against the world, not against Claude.

None of this is a reason to avoid Claude - it is the map that lets you use it professionally. The studio that knows these six patterns, checks in proportion to the stakes, runs a real checklist, and uses Claude as a flagging second reader while keeping a human as the judge, gets the speed without shipping the mistakes. That is the whole of the studio system, and the whole of this course: Claude does the fast, well-read, first-draft work; you supply the truth-checking, the decisions and the accountability - with your scrutiny scaled to the stakes.