Task statement 2.1 · Output evaluation and validation

2.1 · Accuracy and Completeness

In Wrenfield Housing, the course example, the operations team asks Claude to draft its quarterly performance summary for the October board meeting. The draft reads well, but one figure in it is wrong and a section the request asked for is missing. A quick read-through catches neither before the board sees the draft.

Why it's tested

The tempting habit is to read a draft through once and pass it on if it sounds right. Claude's drafts are fluent whether or not every figure is correct, so a wrong number sits in a well-made sentence and reads like a right one. A missing section is harder still, because it leaves nothing on the page to notice. A read-through only checks the draft against itself, so it can confirm that the draft is clear but not that it is right or whole.

KnowledgeAccuracy, completeness, and what to read against

Evaluating a draft means checking two different things. Accuracy is whether every fact, figure, name and date in the draft is right. A right fact matches the source, the material the draft was built from, and fits what the request meant, such as the right period. Completeness is whether everything the request asked for is in the draft, and nothing important in the source has been left out. Neither question can be answered from the draft alone, because Claude does not mark the figure it misread or the part it skipped. The Claude Help Center says Claude can give incorrect or misleading responses and should not be relied on as a singular source of truth. This rules out any answer that judges a draft by how well it reads, or by asking Claude whether it looks right.

So you read the draft against two things outside it: the brief, meaning the prompt that set out what the draft must contain, and the source. The same four steps work for any draft:

  1. Turn the brief into a checklist: number every section, figure and limit it asked for.
  2. Completeness pass: find each checklist item in the draft, and mark any that is missing, thin or folded into something else.
  3. Accuracy pass: find each figure, name and date of the draft in the source, and compare the value, the unit, the period and the direction of change.
  4. Fix and recheck: correct what you found in one targeted follow-up, then check the changed parts again.

Interactive · one draft, read three ways

Three columns for the quarterly summary: what the brief asked for on the left, what the draft says in the middle, and what the figures sheet says on the right. Arrows labelled asked run from each brief item to the draft; arrows labelled checked in run from each draft figure to the sheet. Matching figures are solid and labelled matches. The damp and mould wait is dashed and labelled does not match. The complaints item has no place in the draft, shown dashed and labelled missing. THE BRIEF ASKED FOR THE DRAFT SAYS THE FIGURES SHEET SAYS 1 Repairs: open jobs 1 Repairs: damp wait 2 Rent collection 3 Complaints ASKED ASKED ASKED ASKED 640 open jobs MATCHES FOUND damp and mould: 14 days DOES NOT MATCH FOUND 98.1% of rent collected MATCHES FOUND MISSING NO COMPLAINTS SECTION CHECKED IN CHECKED IN CHECKED IN open jobs: 640 damp wait: 41 days rent collected: 98.1% complaints: 47, 41 answered on time The brief finds the gap; the sheet finds the wrong figure. Read alone, the draft gives no sign of either problem. Every item is found except complaints; no figure is compared. Three figures match, one does not; the gap is never tested.

Choose how to read the quarterly summary, and see which problem each reading can find.

Read it through: every sentence is clear and every figure looks plausible. Nothing in the draft marks the wrong wait or the absent complaints section.

Against the brief: each numbered item is looked for in the draft. Complaints has nowhere to point, so the gap shows. The figures are found but not yet compared.

Against the sheet: each figure in the draft is looked up. The damp and mould wait says 14 days where the sheet says 41. The complaints figures are never reached, because the draft has none to look up.

Solid and labelled MATCHES or FOUND: the check passed. Dashed and labelled DOES NOT MATCH or MISSING: the check found a problem.

The two passes start from different ends, which is why one cannot stand in for the other. The accuracy pass starts from the draft and works outward to the source, so it can only test what the draft contains. A section that is not there gives it nothing to look up. The completeness pass starts from the brief and works inward, so a gap shows up as an item with nowhere to point. This rules out any answer that checks every figure with care but never compares the draft with the request.

KnowledgeWhat each pass compares

Each pass has its own question, its own starting point and its own blind spot, which this box sets side by side:

Draft → source

Accuracy

Is each figure, name and date right? Compare the value, the unit, the period and the direction: a figure copied correctly from the wrong month is still wrong.

Cannot see: anything the draft leaves out.

Brief → draft

Completeness

Is every requested item there, with enough substance? Then scan the source for anything important the brief did not name.

Cannot see: whether the figures in it are right.

Completeness has a second half, because a brief cannot list everything that matters. Anthropic's Claude 101 course warns that Claude may get the right numbers but miss the overall patterns, and it suggests asking what is missing whenever you compare a result with your own work. If the figures sheet showed the backlog rising for a third quarter, a summary that printed the backlog correctly but never said it was rising would be accurate and still incomplete. A numbered brief makes the first half easy. Anthropic's prompting guide recommends numbered lists when the completeness of steps matters, and the same list becomes your checklist.

SkillWorked example: the board summary that read well

Wrenfield's operations team asked Claude for the July–September performance summary, from the quarter's figures sheet and the complaints summary, neither of which holds tenant names. The brief named three sections: repairs, rent collection and complaints. The operations director read the draft against both:

# the brief as a checklist
1 repairs   2 rent collection   3 complaints   · neutral tone · two pages at most
# the draft, line by line, against the checklist and the figures sheet
Repairs: 640 open jobs                        found · sheet: 640, matches
Damp and mould: median wait 14 days           found · sheet: 41 days, does not match
Rent: 98.1% of rent due collected             found · sheet: 98.1%, matches
(no complaints section)                       missing · sheet: 47 received, 41 on time
# one targeted follow-up
Correct the damp and mould median wait to 41 days, as the figures sheet says.
Add the complaints section from the attached complaints summary.
Change nothing else.

The symptom was a smooth draft with two faults that a read-through would have passed. Fourteen days looked like good news and sat in a correct-looking sentence, and nothing in the draft hinted that complaints were missing. The cause was not a careless prompt, since the brief named every section. The fix was a short follow-up naming exactly what to change and what to keep. The director then rechecked only the damp and mould line and the new complaints section, not the whole summary. A claim the director could not find in either file would be a different problem: spotting invented content is 2.2, and verifying a claim beyond your own files is 2.3.

Asking Claude to check helps, but you still check. Anthropic's guidance suggests asking Claude to review each claim in a draft and find a supporting quote in your documents. It also says such techniques reduce errors without removing them, so critical information still needs your own check.

Ask one question: can you point from every item in the brief to its place in the draft, and from every figure in the draft to its source? Yes → the draft is ready for whoever signs it off. No → fix those points in one targeted follow-up, then recheck what changed.
Takeaway
  • Accuracy is checked from the draft to the source; completeness is checked from the brief to the draft.
  • A fluent draft can hide a wrong figure and a missing section, because neither one leaves a mark on the page.
  • A right figure for the wrong period, or a right number with the trend left out, still fails the check.

check yourself

0 / 3 answered

A procurement officer asks Claude to summarise a 30-page supplier contract on one page, covering the price, the term, the renewal terms and the exit clauses. The summary reads clearly and every figure looks plausible. The director will decide from it, without opening the contract, whether to renew.What is the most effective way for the officer to check the summary before sending it?

Select one.

An HR officer asks Claude to turn a training attendance spreadsheet into a report for the leadership team in three parts: attendance by department, the three departments with the lowest attendance, and courses with waiting lists. The report will set next year's training budget, so every figure in it matters.Which TWO checks would most directly find a wrong figure or a missing part?

Select two.

A finance assistant asks Claude for a monthly spending summary by category from an exported ledger, for a budget committee that uses it to spot unusual spending. Every category they asked for is there, and each total matches the ledger. The ledger also shows travel spending doubled from last month; the summary lists the travel total without comment.What is the most effective next step?

Select one.

sources

  1. Claude Help Center, "Claude is providing incorrect or misleading responses. What's going on?": Claude can give incorrect or misleading responses; do not rely on Claude as a singular source of truth, and scrutinise high-stakes advice. Checked 2026-10-15.
  2. Claude Academy, Claude 101, "Getting better results": compare Claude's results with your own work, asking whether it captures the key information and what is missing; Claude may get the right numbers but miss the overall patterns; verify key facts for high-stakes work. Checked 2026-10-15.
  3. Claude Platform Docs, "Prompting best practices": numbered lists or bullet points when the order or completeness of steps matters. Checked 2026-10-15.
  4. Claude Platform Docs, "Reduce hallucinations": after drafting, have Claude review each claim and find a supporting quote in the documents; these techniques reduce errors but do not eliminate them, so validate critical information. Checked 2026-10-15.
  5. Wrenfield's quarterly performance summary, its figures sheet and complaints summary, the 14-day misprint and the missing complaints section are this course's example.