CCAO-F 2.6 — Choosing formats and curating information

What happens to this output next?

Read now, then done
Inline answerin the conversation
Kept, edited or shared
Artifactstandalone, versioned, shareable
Sorted, filtered or imported
Structured datatable, CSV, JSON, spreadsheet
Needed as an Office file
Created fileExcel, Word, PowerPoint, PDF

The format follows the use. The same content — a comparison of three suppliers — could rightly be any of these, depending on who uses it and how.

Inline: the default for conversation

Most answers belong in the conversation: explanations, quick comparisons, short drafts you will paste elsewhere, answers to “what does this clause mean?”. Inline text is fast to read and fast to follow up on. Its weakness is that it lives in the chat history — hard to find again, awkward to share as a finished piece, and mixed in with everything else that was said.

Format inside an inline answer matters too. Anthropic’s prompting guidance notes that the style of your prompt influences the style of the response, and that it works better to say what you want (“flowing prose paragraphs”) than what you don’t (“no bullet points”). A short table is often the best inline format for comparing a few options; a wall of bullets is rarely the best format for an argument.

Artifacts: when the output is a thing, not an answer

Claude’s help centre describes an artifact as something Claude makes that you would put in front of someone — a document, a deck, a dashboard, a diagram or a small interactive tool — shown in its own window beside the conversation. Claude creates one when the content is significant and self-contained (typically more than about 15 lines), stands on its own without the surrounding chat, and is something you are likely to edit, reuse or come back to.

Artifacts give you	Which matters when
A standalone view, separate from the chat	The output will be read without the conversation around it
Editing, and a version selector between iterations	You will refine it over several rounds and may want an earlier version
Copy and download options	It will be used outside Claude
Sharing by link, with plan-dependent access controls	Colleagues need to see or reuse it
Interactive content such as dashboards or small tools	The reader needs to explore, not just read

Two practical notes from the help centre. Artifacts need the code execution and file creation capability switched on in settings (on Team and Enterprise plans, an owner controls it). And sharing has consequences: when you publish or share, attachments from the original conversation can become visible to viewers, and some artifacts store information that everyone who opens them can see — so check what is in the conversation before you send the link. Sensitive-data handling in general is covered in 6.2.

Structured data: when a spreadsheet or system is the reader

Some outputs are read by a tool before a person: a list to sort in Excel, a table to import into a CRM, fields to paste into a form, a JSON file for a colleague’s system. Prose is the wrong format here, even if it is well written, because someone has to re-type it. The right move is to ask for structure explicitly, and to specify it precisely. Anthropic’s consistency guidance puts specifying the exact output format first among its techniques, and recommends showing an example of the format you want.

Asking for data you can actually use

Produces prosetext

Summarise the pricing and terms in
these 12 supplier quotes.

Produces datatext

From the 12 attached supplier quotes,
build a table with one row per supplier
and these columns, in this order:

supplier | unit_price_gbp | min_order |
lead_time_days | payment_terms | source_page

- Prices in GBP, numbers only, no symbols
- If a value is not in the quote, write
  NOT_FOUND - do not estimate
- Give me the result as a CSV file

The right-hand request fixes the columns, units, one-row-per-item rule and what to do with missing values — which is what makes the result sortable and importable without clean-up.

With code execution and file creation on, Claude can also produce real files — Excel spreadsheets with working formulas, Word documents, PowerPoint decks and PDFs — rather than text you paste into them. That is the right choice when the recipient works in those tools. Structured outputs still need the checks from 2.1 and 2.3: a tidy table with a wrong number is still wrong, and “NOT_FOUND” values are gaps to chase, not to hide.

Worked example · One HR question, three formats

Keisha, an HR business partner, is updating the company’s parental leave policy. In one afternoon she uses Claude for three different things.

A manager asks whether the new policy changes anything for someone already on leave. Keisha asks Claude against the policy document and reads the answer inline. One question, read once — no artifact needed.
The new-starter guide to the policy will be edited by the legal team, reused every quarter and shared with managers. She has Claude draft it as an artifact, iterates on it with version history, and shares the link with legal.
The HR system needs each leave type as a row: type, weeks, pay rate, eligibility rule, effective date. She asks for exactly those columns, one row per leave type, as a CSV file, and checks every row against the signed-off policy before import.

Takeaway: Same source, same afternoon, three formats — each chosen by what the recipient would do with it next.

Organising and curating information

Curating is deciding what the reader gets: what to include, what to leave out, how to group it and in what order. It is where Claude saves the most time — turning fifteen sources into a two-page brief — and where the most can quietly go wrong, because every merge is a chance to lose a caveat or the source of a claim.

Organise by the reader’s question, not by source. “What competitors charge” is more useful than “what Source 1 said, then Source 2”.
Rank. Put what changes the decision first; move background to the end or an appendix.
Keep attribution attached. Each figure or claim should carry its source, so curation does not erase the trail back to the evidence.
Flag conflicts instead of blending them. If two sources disagree, say so and show both; an average of two contradictory figures is a figure nobody reported.
Say what was left out. A line on what was excluded, and why, keeps the curation honest.

How you give Claude the material helps too. Anthropic’s long-context guidance recommends putting long documents at the top of the prompt with the question at the end, and wrapping each document with its name and source so Claude can tell them apart — which also makes it easier for the output to cite them.

Worked example · A competitor brief from fifteen sources

Tomas, in product marketing at a payments company, uses Research and a folder of analyst notes to prepare a brief on three competitors for the sales leadership meeting.

He sets the structure first: pricing, target customers, recent launches, and where sales teams are losing deals — the questions leadership actually asks.
Claude’s first draft is organised by source, fifteen short summaries in a row. Accurate, but nobody could use it in a meeting.
He asks for a reorganisation by his four questions, with the source after each point and any disagreements between sources shown side by side.
One conflict surfaces: two sources give different prices for the same competitor plan. The brief shows both with dates; he checks the competitor’s pricing page before the meeting.
Format: the brief goes out as an artifact for leadership to annotate, and the pricing comparison as a small spreadsheet the sales team can filter.

Takeaway: Curation turned a pile of accurate summaries into something usable — by reorganising around the reader’s questions while keeping every claim tied to its source.

Is this curated brief ready?

Passes: Organised by the reader’s questions
Passes: Most decision-relevant points first
Check: Every figure carries its sourcetwo launch dates unattributed
Passes: Conflicting sources flagged, not mergedpricing conflict shown with dates
Missing: Exclusions statedadd a line on what was left out
Passes: Format fits the useartifact brief plus filterable spreadsheet

Curation can fail even when every sentence is accurate: the organisation, the attribution and the handling of conflicts are what these checks test.

Traps the wrong answers are built from

Tempting but wrong	Do this instead
An artifact for a one-off answer	Keep quick, read-once answers inline.
Prose for data someone must sort or import	Ask for a table, CSV or spreadsheet with named columns and a missing-value rule.
Vague format requests like “put it in a table”	Specify columns, order, units and one row per item — and show an example.
Summaries organised source by source	Organise by the reader’s questions and rank by importance.
Blending conflicting sources into one figure	Show both with their sources and dates, and resolve it at the source.

You should now be able to

Choose inline, artifact, structured data or a created file from how the output will be used next.
Apply the help centre’s artifact criteria: self-contained, substantial, likely to be edited, reused or shared.
Specify structured outputs precisely: columns, order, units, one row per item, missing-value rule.
Curate many sources into a brief organised by the reader’s questions, with attribution kept.
Flag conflicts and exclusions instead of hiding them.
Check what a shared artifact exposes before sharing it.

Choosing formats and curating information