Designing · The devil is in the TL.DR
What a summary leaves out
Aggressive summarisation hides assumptions and tradeoffs users must see. Summaries must preserve material conflicts, sources, and approval needs while still reading well.
10 min read
Cover for What a summary leaves outThe Meal Companion presents a weekly plan summary: seven balanced dinners, spinach used, allergies respected. Clean. Fast. Trustworthy until the household discovers Tuesday was compressed because of a calendar conflict they never saw, or that "allergies respected" hid an unendorsed inference from a chat line rather than an enforcement result from a deterministic check. Summarisation is where trust is won or lost. A summary that erases busy-day reasoning, allergy sources, or contested trade-offs produces passive acceptance, not informed approval.
Helpful context: Confidence theatre and contestability explain why compression needs grounding. Visible thinking is the design discipline for what must survive summarisation.
What must survive summarisation
For a weekly meal plan summary to support informed approval, certain distinctions must survive compression. Material assumptions must remain visible: busy evenings mean at most twenty-five minutes active effort. Constraint sources must be cited: cashew allergy comes from endorsed knowledge rather than a chat rumour. Active conflicts must surface: spinach deadline versus guest night creates a tension the household may want to resolve differently. And approval needs must be explicit: grocery purchase requires human sign-off, medical terms defer to a clinician.
Omit any of these and the household approves a plan they cannot contest. The summary saved words, but it did not remove cognitive effort. It moved effort to a costlier moment: shopping time, dinner time, or the support conversation where the household says "why did you do that?" and the product cannot answer.
Three kinds of compression
The distinction between compression strategies matters because each carries different risks. Extractive summary selects sentences from the source material, making hallucination harder but synthesis weaker. Abstractive summary generates new prose that reads well but can assert claims the source never made, such as "allergies respected" when no enforcement ran. Grounded compression links each consequential claim to a source or assumption, making omission accountable rather than invisible.
The product instinct to "make it simple" often defaults to abstractive summary because it reads best. The risk is that fluent prose can invent constraints, merge distinct concerns, or assert outcomes without evidence. Grounded compression is harder to write but safer to trust because every material claim traces to a check result, a calendar event, or an endorsed record. Measure effort removed, not words saved. Compression fails when people must reopen the full source to verify claims the summary invented.
Where compression helps and where it harms
The answer differs by Collaboration Contract. A person meant to approve needs enough structure to reject. A person meant to explore may need more room than a summary allows. Compression that hides a contested trade-off is passive acceptance theatre dressed as brevity.
The design process for accountable summaries begins with naming the decision the summary is meant to support. The weekly plan summary supports a planning approval decision. That decision requires the household to assess whether constraints were respected, whether conflicts were resolved acceptably, and whether any actions still need human sign-off. The distinctions that must survive compression follow directly from the decision type. Citations or source badges attach to each consequential claim. The test of success is whether people still reopen the full source to verify: if they do frequently, the summary hid too much.
Designing accountable summaries
Expandable panels beat wall-of-text and stripped-down summaries equally. Each hidden assumption gets a "why this" link that expands into source and reasoning. Approval summaries list what was checked and what passed, not outcomes alone. The weekly meal summary surfaces one line per material constraint with a source badge: "Busy Tuesday" from calendar, "Allergy" from endorsed knowledge, "Purchase" pending approval.
The shape of a good weekly summary leads with a decision-ready headline, lists material constraints with sources, flags conflicts and approval gaps, and offers expand for each day when the household wants meal-level detail. Keep recommendation patches visible when critique found weaknesses, because hiding the fix is as misleading as hiding the problem.
Summaries function as UI for approval mode. Treating them like marketing copy that rewards word-count reduction produces fluent omission. Treating them like legal cover sheets that reward accountable claims produces trustworthy compression. The analogy is not accidental: regulated approval workflows already know this pattern, and meal planning inherits the same requirement as stakes rise with purchase and medical deferral.
The Meal Companion bad summary
"Seven balanced dinners, spinach used, allergies respected" reads well. It hides that Tuesday was chosen because calendar marked busy, that cashew avoidance came from endorsed knowledge and not chat, that grocery purchase still needs approval, and that guest Thursday may conflict with spinach timing. A household approves in haste and discovers problems at shopping time. The summary saved words. It did not remove cognitive effort. It moved effort to a costlier moment.
The same pattern repeats in enterprise domains with higher consequence. An invoice summary that says "all checks passed" without listing which checks ran on which line items produces the same approval failure at audit time. The risk scales with consequence, but the architectural mistake is identical: abstractive compression without grounding.
Evaluating summary quality
Two metrics reveal whether summaries are working. Reopen rate measures how often users open the full source material after approving the summary. High reopen means compression hid too much because the summary did not answer the questions it should have. Override rate measures how often users change their mind after initial acceptance. High override means hidden assumptions surfaced too late.
Offline graders can check whether mandatory fields survived compression on held-out plans. A rubric lists which constraints must appear (allergy always material, minor garnish choice may compress), which sources must be cited, and which conflicts must remain visible. Rubrics beat word count targets because word count rewards fluent omission while rubrics reward accountable retention.
Thoughtware: Designing in the Intelligence Age · Ch. 12Summarisation is where trust is won or lost, so make omission accountable.
Expandable compression as middle path
The wall-of-text problem is real. Households do not want to read ten paragraphs about Tuesday's meal selection. The TL.DR-with-hidden-assumptions problem is equally real. Grounded compression with expandable detail is the middle path: default collapsed summary with one line per material constraint, expand per constraint for source and reasoning, and approve with confidence that the right information was available at the right depth.
The expand affordance changes the design contract. A collapsed summary no longer needs to contain everything because expansion is one tap away. The constraint becomes: collapsed copy must contain enough for the household to know whether to expand, and expansion must answer the question completely without further drilling. Two levels of depth serve most approval decisions. Three levels indicate the summary architecture is trying to compress too many concerns into one surface.
Summary versioning and plan evolution
After a patch changes Tuesday's meal, the summary should update with a change log line: "Tuesday swapped for quick pasta, busy rule applied." Summaries track plan evolution across patches, not initial state alone. A version mismatch between summary and plan is a trust bug because the household reads stale grounds while the underlying plan has moved.
The same principle applies to email and notification summaries. Email strips expand affordances, so material constraints should survive in plain text or the email should link to a grounded view with full expand capability. Smart speaker readout strips everything visual, so voice summary scripts must include mandatory constraint lines aloud. Multimodal summary risk is often missed in visual-first design teams.
Common failures
The most frequent failure is abstractive summaries that invent claims: "allergies respected" when enforcement did not run. A close second is TL.DR obsession with word count rather than effort removed. Approval summaries that list passes without listing checks performed produce false confidence at audit time. Embedding merge on paraphrase can collapse distinct constraints into one summary line silently. And summaries that assert outcomes without linked checks fail the moment a regulator, auditor, or concerned household member asks "how do you know?"
These failures share a common root: treating summarisation as a language task (make it shorter) rather than a governance task (make omission accountable). The architectural response is grounded compression with mandatory survival fields, source badges, and expand paths. Train models and templates on grounded compression rubrics, not on word count targets alone.
What to do next
The path toward accountable summaries starts with the product's primary summary view. Highlight every material assumption it currently hides. Each hidden assumption becomes a candidate for an expandable "why this" panel with source badge. Define a materiality table that specifies which constraints always survive compression (allergy, approval needs, active conflicts) and which may compress safely (garnish preference, variety reasoning). Apply that table as a rubric in offline eval so summary quality becomes measurable rather than subjective.
See recommend, don't only respond and contestability.
Read next: Visible thinking.