PTN—18
Progressive Disclosure
Lead with the answer at skimmable depth and put reasoning, sources, and detail one deliberate step away.
Models are verbose by default, and verbosity taxes the reader on every single response. Progressive disclosure is the discipline of layering: the answer first, at the depth most people need, with reasoning chains, full source excerpts, methodology, and edge cases behind explicit expansion points. The person who wants the number gets the number; the person who wants to audit it can, without either paying for the other’s preference.
This is an information hierarchy problem before it is a UI problem. The collapsed layer has to be genuinely sufficient — a real answer, not a teaser engineered to force a click. And the expanded layer has to be genuinely deeper — the actual chain of reasoning or the actual source text, not a restatement. When both layers hold, disclosure becomes a trust instrument: the detail is verifiable precisely because it is optional.
Reasoning models made this pattern load-bearing. Extended thinking can run to thousands of words that almost nobody should read linearly, yet hiding it entirely would violate the contract of showing your work. The shipped consensus — thought process collapsed under a labeled toggle, answer front and center — is progressive disclosure doing exactly its job: full transparency at zero default cost.
The tradeoff is editorial nerve. Every layering decision is a claim about what most readers need, and teams that lack the nerve hedge by promoting everything to the surface — which is just verbosity with extra chrome. Maturity means instrumenting the toggles: a disclosure opened on most responses belongs in the default layer; one that is never opened should justify its existence. Perplexity’s evolution is the model case — citations moved from footnote lists toward inline markers as usage proved verification was a primary behavior, not an expert one. The layers should follow the readers, not the org chart.
When to use
- Responses that carry reasoning, sources, or caveats most readers will not need most of the time.
- Reasoning-model outputs where the thinking trace is long but must remain inspectable.
- Mixed audiences — the same response serves skimmers and auditors.
- Dense outputs like analyses or comparisons where a summary layer genuinely stands alone.
Design considerations
- 01Make the collapsed layer a complete answer — if expanding is mandatory to understand it, the layering failed.
- 02Label expansion points with what they contain: “Show reasoning”, “View 12 sources” — never a bare chevron.
- 03Keep the toggle adjacent to the claim it substantiates, not pooled at the end of the response.
- 04Expand in place without losing scroll position or collapsing on the next streamed token.
- 05Cap the layers at two, maybe three; disclosure trees deeper than that are navigation, not reading.
- 06Persist a user’s expansion preference within a session instead of resetting every turn.
Pitfalls
- ✕Burying the safety-relevant caveat in the collapsed layer, where disclosure quietly becomes concealment.
- ✕Teaser summaries that force an expansion click on every response — disclosure as engagement farming.
- ✕Expanded content that merely restates the summary in more words, teaching people the toggle is worthless.
- ✕Collapse state that resets on every new message, so readers re-expand the same sections all session.
In the wild
- Claude
- Extended thinking collapses the reasoning trace under a toggle, keeping the answer primary while the full chain stays inspectable.
- Perplexity
- Answers lead with synthesized prose; numbered citations expand to source context only when invoked.
- Gemini
- Thinking-model responses tuck the reasoning into a collapsible section above the final answer.
- GitHub Copilot
- Inline completions show a single suggestion; alternatives and explanations sit behind explicit controls rather than crowding the editor.