Task statement 2.5 · Output evaluation and validation

2.5 · Adapting Outputs for the Audience

In Wrenfield Housing, the course example, the operations team has a plan for the repairs backlog, and three groups must hear about it: tenants on the two estates with the most damp and mould jobs, the board, and the contractor. One update written for all three suits none of them.

Why it's tested

The tempting habit is to send one accurate update to everyone, or to ask Claude to "make it simpler" and accept the first version. Accuracy is not enough, because each reader came for something different: a tenant needs to know what to do, a board member needs the decision, and a contractor needs tasks and dates. Rewriting every version by hand fails too, because it throws away Claude's speed. A hand edit that Claude never sees is undone by the next request.

KnowledgeWhat changes from one reader to the next

The intended audience is the person who will read an output and act on it. An output fits when that reader can find what they need and act on it. Four things usually change from one reader to the next: the reading level, meaning how hard the words and sentences are; the length; what the reader must do; and the tone. The facts do not change, so you adapt every version from one checked source. This rules out any answer that judges an output on accuracy alone, or that treats "simpler" as the fix for every reader.

The exam guide names four ways to work on an output for its reader, and each one suits a different moment:

  1. Edit: fix a small, known problem in the text yourself, such as a wrong date or a name.
  2. Adapt: ask Claude for a version for one reader, naming the reading level, the length, what they must do and the tone.
  3. Refine: improve that version in targeted passes, one quality per pass.
  4. Compare: when two versions compete, read them side by side against written checks for that reader, and pick one.

Interactive · one update, adapted for three readers

One backlog update at the top, with an arrow labelled adapt down to each of three versions: tenants, board and contractor. Each version reads top to bottom: what it leads with, what it keeps, what it leaves out, and its length and tone. The chosen version is solid and labelled chosen; the other two are dashed and faint. ONE CHECKED SOURCE Backlog plan: 640 open jobs · damp and mould 112 jobs, median 41 days Second contractor team from 3 November · £48,000 · target 20 days by end of January ADAPT TENANTS TENANTS · CHOSEN LEADS WITH what to do: report damp, keep the visit we book KEEPS visits from 3 November, rebook via the repairs line LEAVES OUT the cost, the job counts, the contract terms LENGTH · TONE about 150 words, plain, warm ADAPT BOARD BOARD · CHOSEN LEADS WITH the decision: approve £48,000 for 12 weeks KEEPS 640 jobs · 112 damp and mould · 41 → 20 days LEAVES OUT how tenants book visits, the contractor's rota LENGTH · TONE one page, neutral, formal ADAPT CONTRACTOR CONTRACTOR · CHOSEN LEADS WITH the tasks: survey all 112 jobs by 28 November KEEPS start 3 November, book in date order, report Fridays LEAVES OUT the board's reasons, the tenant wording LENGTH · TONE short, direct, every task dated Same facts in every version: the reader decides what leads, what stays and how it sounds.

Choose the reader for the backlog plan.

Tenants must act, so their version opens with what to do and drops the cost and job counts.

The board must decide, so its version opens with the decision and keeps the figures behind it.

The contractor must deliver, so its version opens with the tasks and gives each one a date.

Solid and labelled CHOSEN: the reader you picked. Dashed and faint: the other two versions.

Adapting works best when the prompt names the reader and the four changes. A Claude Academy example asks for versions of one handout with "same concepts, different vocabulary and sentence length". It also advises asking for a list of what Claude changed and what it left out whenever it rewrites, simplifies or adapts your text. That list is what you review. For tone, Anthropic's prompting guide advises telling Claude what to do instead of what not to do, so "write in short sentences a tenant can act on" steers better than "no jargon".

KnowledgeEdit it yourself, or ask Claude to change it

An edit is a change you make to the text yourself; a request asks Claude to make the change. Edit when the change is small and you already know the answer, such as a wrong figure or a name, because typing it is faster and nothing else moves. Ask Claude when the change runs through the whole text, such as the reading level, the length or the tone, because Claude reworks every sentence faster than you can. Mixing the two has one trap. Claude sees only the conversation, so a correction made in your own document is invisible to it. The next request starts from its uncorrected draft unless you paste your version back:

Small and known

Edit it yourself

A wrong date, figure or name: you know the fix, and nothing else should move.

Then: paste the edited text back before the next request.

Runs through the text

Ask Claude to change it

The reading level, the length or the tone: every sentence moves, so Claude does it faster.

Then: ask what it changed and what it left out, and check that list.

Refining is improving one version in targeted passes, each aimed at one quality for that reader: the facts against the source, then the reader's action, then length and tone. It reuses the loop from 1.3: written checks and one change per turn. The draft is already close, so each pass names the part to change. Claude Academy rates "Cut the first two paragraphs and make the conclusion more action-oriented" above "Make it shorter".

Comparing is reading two versions side by side against the same written checks for the reader. You can ask Claude for versions with different approaches in one turn, as a Claude Academy example does with "three versions: safe, debatable, and contrarian". Write the checks from the reader's needs before you read either version, because a version that reads better to you may still fail the reader. Pick the one that passes more checks, then refine its gap rather than merging the two.

SkillWorked example: two tenant versions, one choice

Wrenfield's plan adds a second damp and mould contractor team from 3 November, to cut the median wait for damp and mould jobs from 41 days to 20 by the end of January. The team asked Claude for two tenant versions, one formal and one plain, and wrote four checks before reading either:

# checks, written before reading
1 what to do is in the first three lines   2 no costs or job counts
3 short, plain sentences                   4 visits start 3 November
# version A, 260 words
"Following a review of our backlog of 640 open jobs, the board has
approved £48,000 for additional damp and mould remediation …"
→ passes check 4 only: what to do is in the last paragraph
# version B, 140 words
"Damp or mould at home? Report it on the repairs line. We will book
a visit, and you can rebook if you cannot be in …"
→ passes checks 1–3; no start date, so refine one pass:
Add that visits start on 3 November. Keep everything else as it is.

Version A was accurate but written for the board: it opened with figures a tenant cannot act on and used words such as "remediation". The four checks showed this where "A sounds more professional" would have hidden it. The team picked B because it passed more checks. They closed B's one gap with a targeted pass, because merging in A's opening would have brought the figures back. In the board version, Claude wrote the cost as £84,000, so the operations team corrected it by hand and pasted the corrected text into the chat before asking for anything else. Before they go out, the communications lead checks the tenant version and the operations director checks the board version, as Wrenfield's AI policy requires.

Ask one question: what does this output need before its reader can use it?
A small fix you already knowEdit it yourself, and give Claude your version.
A version for a new readerAdapt from the checked source: name the reading level, length, action and tone.
A choice between two versionsCompare them on checks written from the reader's needs, then refine the one you pick.
Takeaway
  • The facts stay fixed across readers; the reading level, length, required action and tone change with each one.
  • Claude never sees changes made outside the chat, so a hand edit must be pasted back before the next request.
  • Choose between versions by the checks each one passes for its reader, not by which one reads better to you.

check yourself

0 / 3 answered

An IT project lead has a two-page update on moving staff email to a new system, written for the IT steering group. The lead now needs a version for all 400 staff, who must each reset their password between 2 and 6 June and will mostly read it on their phones.What is the most effective request to send Claude?

Select one.

A charity's fundraising officer asked Claude for two versions of an appeal email to past donors, one built around a story and one around figures. The officer must choose one by Friday, and their manager has asked them to say why that version was chosen.Which TWO steps would best help the officer choose?

Select two.

A school business manager asked Claude to turn a governors' budget report into a letter for parents. The manager corrected two wrong dates in the letter by hand in their own document, then asked Claude, in the same chat, to make the letter shorter. The shorter letter has the two wrong dates again.What is the most likely root cause?

Select one.

sources

  1. Claude Platform Docs, "Prompting best practices": control the format of responses by telling Claude what to do instead of what not to do. Checked 2026-10-15.
  2. Claude Academy, "Adapt a standard textbook page to every reading level": versions with the same concepts but different vocabulary and sentence length; when Claude rewrites, simplifies or adapts, ask for a list of what it changed and what it left out, and review that list. Checked 2026-10-15.
  3. Claude Academy, "Adapt content across platforms": rewrite copy that is too technical for everyday users; review the output and tailor tone and complexity; ask for "three versions: safe, debatable, and contrarian". Checked 2026-10-15.
  4. Claude Academy, "Claude 101: Getting better results": describe the tone in plain language; be explicit about length; specific feedback ("Cut the first two paragraphs and make the conclusion more action-oriented") beats "Make it shorter". Checked 2026-10-15.
  5. Wrenfield's backlog plan, its three versions, the four tenant checks and the £84,000 slip are this course's example.