The Clinic

Class: wrong shape. Every fact is right. The reasoning is right. It is the wrong artifact — a post-incident report for engineers, where a status page update is a short customer-facing paragraph with no component names, no version numbers and no root-cause narrative.

The repair — one line, or better, one example. Naming the artifact precisely does most of the work:

"Status page update, not an incident report. Sixty words, one paragraph, plain customer language. No internal component or version names, nothing about root cause. Say: what was affected, the window, that it's resolved, whether anyone needs to do anything. Here's the one we posted last time — match it: [paste]."

Compare against your run on this: did you exclude the SLA line and the "nobody has told support" line? The notes contain three things that must not appear on a public status page, and the original brief never said so.

The arguable alternative. You could call this under-specified, on the grounds that the non-goals were never stated. It's a reasonable read, and here is why wrong shape is the better one: naming the artifact fixes the exclusions for free. "Sixty-word customer-facing status update" already forbids the version numbers and the SLA discussion. When the right container implies the constraints, reach for the container.

Class: insufficiently decomposed. It did all of it, at one pass, and the effort is spread evenly across parts that carry wildly different marks. Two paragraphs on Innovation, which is worth 5. Three sentences on Mobilisation, which is worth 20. Social value, worth 15, appears once in the opening flourish. The six sub-criteria are not answered separately, so under rule 4.4 several of them are unaddressed, and unaddressed scores 0, and 0 is disqualifying.

The repair is a sequence, not a sentence. That is the honest shape of this one — it is the case in this clinic without a one-line fix:

"Step 1, and stop after it: read the ITT and give me a table of every scored sub-criterion, its marks, its word limit, and any disqualifying condition. Nothing else yet."
Then, one at a time: "Draft sub-criterion (a), mobilisation and transition, in 500 words. Answer only that sub-criterion. Here are the three transition projects we can cite: […]. Show me the evidence you're relying on for each claim before you draft it."
Then assemble, and check each answer against its own word limit last.

Compare against your run on this: did your step 1 surface rule 4.4? The disqualifier is the highest-stakes item in the extract and it is buried three clauses into a numbered paragraph. Extracting the criteria as a discrete step is what makes it visible; asking for the whole bid at once is what buries it.

The tell to learn from this case. Insufficiently decomposed output is uniformly shallow rather than wrong. Nothing in that response is false, and that is exactly why it is easy to accept. If you catch yourself thinking "this is fine, it just needs more detail everywhere" — that is the tell.

Class: fabrication. Wrong specifics inside an otherwise sound answer. Note what makes it dangerous: the organisations are real, the studies they publish are real, the framing is exactly right for the proposal, and the figures are formatted to two significant figures with page references. Nothing on the surface separates this from a well-sourced paragraph. It was asked with nothing attached and search off, so it was generated rather than retrieved — and that condition is the single best predictor you have.

The repair — one line, and you have two routes. Ground it:

"Use web search. For every figure, give me the URL and quote the sentence it came from. If you can't find a source for a figure, say so and leave the figure out rather than estimating it."

Or scope it, which is the better move for this particular proposal:

"Attached are our last eleven engagement records. Compute our own median time from signed letter to first deliverable, show the working, and list any record you had to exclude and why. Use nothing outside the attachment."

Compare against your run on this: whatever came back, check one figure against its stated source before you read any further. That habit — not a belief about how often this happens — is what Lesson 5 is built on. Record what you found, including "the sources checked out."

Class: under-specified. It is accurate. It is well organised. Every theme is genuinely in the notes. It is also what you get when nothing in the request narrows the target — five balanced themes, no ordering by importance, no recommendation, and a closing sentence that recommends more research, which is the single most typical way a research summary can end.

The repair — one line, carrying four components. The decision it feeds, the audience, the shape, and the non-goals:

"This goes to three people who are deciding on Friday whether to keep the Saturday slot. One page. Lead with a recommendation, then the strongest evidence for it and the strongest evidence against it, quoting the respondent. Then say what would change your answer. Don't restate the method, and ignore anything not about delivery."

Compare against your run on this: two things to check. Did yours exclude R5 and R14 — the packaging monologue and the refund complaint — or did they reappear as an "adjacent feedback" theme? And did it use R11, the business account, as the real complication? R11 doesn't want Saturday and wants something else entirely, which is the one finding in these notes that could change what the decision even is.

Not to be confused with wrong shape. The original output has no stated length or format problem to fix — it is a perfectly respectable summary. Reformat it and you have a well-formatted summary that still won't get anyone to a decision on Friday. The fault is in what it selected, not in how it was presented.

Class: missing information. And this is the case in the clinic that cannot be repaired by any prompt.

Read the output again and notice what it has actually done. It has answered competently about a senior engineer at a mid-size software company. It is not about Devan, because nothing about Devan's actual position reached it. The 9% is an assumption — stated once, in a subordinate clause, and then used as a fact for the rest of the answer, including inside the draft you would have sent.

The repair for missing information is to supply what it couldn't have had. Here, work out what that is: Devan's current salary, the band floor and midpoint, and what the two peers are on. Those numbers exist in a payroll system, in an HR document you have not been given, and in two salaries you were told in confidence or never told at all. There is no wording that reaches any of it. Not a better prompt, not a bigger model, not a different setting. The information is not on the web, not in your files, and not in the conversation.

The correct move is to leave the chat and go and get the figures — from your HR partner, or from the compensation portal you do have access to. That is a completed diagnosis, not a failure to prompt well.

If you classified this as fabrication, because of the 9%: you are reading the symptom rather than the cause, and the output did flag its own assumption once before building on it. It doesn't matter much — both routes end at the same wall, because the repair for fabrication is also "give it a source," and there is no source you can reach. Either way the answer is stop.

The fixable task next door. This is worth doing, and it is not the same thing as fixing this case: "Draft the evidence and scope section of a band-change case for Devan. Leave every figure as a bracketed placeholder I will fill in from payroll. Do not estimate any number, and list at the end exactly which figures I need to collect." That is a real repair to a different request. Recognising the difference between repairing your request and replacing it is most of what calibrated persistence means.

Class: lost the thread. Three separate signatures, all in one paragraph: a constraint from turn 1 broken (Castleford named), a second one broken (03/03/27), and a figure that contradicts one agreed in the same conversation at turn 9 — 41 minutes has become "under 30".

The repair: a fresh chat with the brief restated. Not a scolding reply in the same thread. Carry forward four things and nothing else:

"Northwind's response to the Harbourview tender, mobilisation section. Three fixed rules: never name Castleford or any incumbent in writing; dates as '3 March' not 03/03; every response-time claim comes from the 2026 SLA report and nothing else — median first response is 41 minutes (table 6), and no other figure may be used. Here is the current draft: [paste turn 22]. Here is what I want changed: […]."

Compare against your run on this: did you carry the draft across as well as the rules? People restate the constraints and leave the best work behind, then spend three turns regenerating it. The reset is meant to drop the forty turns of detour, not the output they produced.

Why not just remind it in the thread? Because it is cheap and it often works, so it is worth one try — but the same twenty-two turns are still in front of it, including a redraft that says "under 30 minutes", so the same fault has somewhere to come back from. The reset is the move that removes the cause.

Not fabrication. "Under 30 minutes" is a wrong specific, which is tempting. The tell is that it contradicts a figure agreed earlier in this same conversation, with a source attached. A number that drifts away from one you already established is a thread problem. A number that appears from nowhere, with nothing established, is a grounding problem.