Task statement 2.6 · Output evaluation and validation
2.6 · Curating Information and Output Formats
In Wrenfield Housing, the course example, the operations team has typed up 86 anonymised comments from the spring tenants' forum. A colleague, the team's tracker and the board each need something different from them. When every result comes back in one shape, someone reshapes it by hand, and errors slip in.
share
The tempting habit is to accept whatever shape the first reply has, because the content looks right. A long chat reply is easy to read but hard to reuse, so its counts get copied into a spreadsheet by hand. One mistyped number then sits in the tracker as if it were checked. Asking for everything Claude found fails too, because it buries the points the reader needed. The right shape depends on what the reader does with the result next.
KnowledgeOrganising and curating before you choose a format
Organising information means grouping related points and ordering them for the reader, such as the most-raised theme first. Curating means choosing what stays. You cut what this reader does not need, and you keep each claim's source next to it, so the person who checks the result can trace every point back. A request that names none of these leaves each choice to Claude. Anthropic's prompting guide asks you to be specific about the output format. For long material, it suggests asking Claude to quote the relevant parts first, which keeps the evidence beside each point. This rules out treating a well-written reply as finished when nobody can trace its numbers.
Before you ask for the result, settle these in order:
- Group: name the themes or categories, or ask Claude to propose them and say what counts as one.
- Order: sort by what the reader cares about most, such as the count or the urgency.
- Cut: say what this reader does not need, such as individual comments for a board.
- Keep sources: ask for the note, row or page each claim rests on.
- Choose the format: by what happens next. An artifact is a document Claude builds beside the chat. Structured data is rows and columns.
Interactive · one set of forum notes, three formats
Choose a format for the forum themes, starting with the one that went wrong.
Long chat reply: the tracker needed rows but got 900 words of paragraphs, so an officer copied the counts by hand and typed 18 as 13.
Inline answer: one fact, read once, so a two-line reply in the chat is the whole job.
Structured data: the tracker sorts by column, so the themes come back as a table and nobody retypes a count.
Artifact: the board summary will be edited, checked and exported, so it lives as a document beside the chat.
KnowledgeInline, artifact or structured data
The exam guide names three output formats. Inline means the reply in the chat itself, which suits a short answer that someone reads once. An artifact is a standalone piece of content in a panel beside the chat, such as a document, a deck or a dashboard. Claude makes one when the content is significant and self-contained, typically over 15 lines, and likely to be edited or reused. You can also ask for one by name. You change it by asking Claude, by typing into the document, or by highlighting text and choosing "Edit with Claude". Artifacts are saved in the Artifacts tab, and a document exports to Word, PDF or Google Docs. They start private to you, and on a Team plan sharing stays inside the organisation by default.
Structured data is information in fixed fields, such as rows and columns, so that a person or a tool can sort, filter or count it without retyping. It can be a table in the reply, or an Excel file that Claude creates for you to download. Claude's help article warns that a created file may need review and refining, so name the columns you want. Here are the forum themes in each format:
A reply in the chat, for a quick answer read once: "Yes: 14 of the 86 comments raised the repairs line's new hours, mostly asking for Saturday opening."
Wrong whenThe result will be copied, edited or reused.
A table with fixed columns, or an Excel file: Theme, Count, Comment IDs, Example comment.
Wrong whenA reader needs an argument or a story, not rows.
A one-page document beside the chat, edited over several turns and exported for others.
Wrong whenThe answer is a line or two that nobody keeps.
SkillWorked example: the forum themes that were typed in by hand
Wrenfield's operations team typed up the spring tenants' forum notes as 86 comments, numbered C01 to C86, with names and addresses removed first, as the AI policy requires. The first prompt asked an open question, so the answer came back as the only shape that question implies, a long reply in the chat:
# before [86 anonymised comments, C01–C86] What did tenants say at the forum? → 900 words of paragraphs; counts copied into the tracker by hand, damp and mould typed as 13 instead of 18 # after: the tracker Group the comments into themes, most-raised first. Give an Excel file with the columns Theme, Count, Comment IDs, Example comment. → Repairs speed 31 · Damp and mould 18 · Repairs line hours 14 · Communal cleaning 12 · Other 11, each row with its comment IDs # after: the board pack From that table, write a one-page summary for the board as an artifact: the top three themes with counts, and what we are doing about each. Leave out individual comments. → a one-page document, edited, then exported to Word
The cause was a format chosen by default, not a weak answer. Every count in the long reply was right, but the tracker needed rows, so the hand copy brought in the one error. The fix organises first, with themes in count order and comment IDs beside each count. It then gives each reader a format: a table for the tracker, an artifact for the board with the individual comments cut, and a two-line inline answer when a housing officer asks about the repairs line's hours. The operations director checks the board summary first, as the AI policy requires.
More precise than the guide. The guide lists three formats, but in the apps, artifacts and created files such as Excel spreadsheets both need code execution, a capability that lets Claude run code to build files. It is turned on together with file creation. On Team and Enterprise plans it sits in the organisation's Capabilities settings, which an owner controls (checked 2026-10-15). Exam answer: inline for a quick answer, an artifact for content that is edited, reused or shared, and structured data for content a tool or person will sort or count.
| Someone reads it once and moves on | Inline: a short reply in the chat. |
| It is edited over several turns, kept or shared | Artifact: a document beside the chat, exported when it is done. |
| A tool or a person sorts, filters or counts it | Structured data: a table or a file with named columns. |
- Organise for the reader and curate for the checker: group and order the points, cut what this reader does not need, and keep each claim's source beside it.
- A result reshaped by hand picks up errors that the original never had.
- Claude creates an artifact for substantial, self-contained content you will edit or reuse, and you can ask for one by name.