Product design
What should the product be, and how should it work?
Choices about what a product does and how it behaves: its core interaction, its defaults, what it ranks or recommends.
29
48%
39%
13%
What the cases suggest
Product design decisions by young companies mostly paid off in this record, while those by established leaders were more often mixed or too early to judge. Many of the leaders’ design decisions are recent, concerning AI features and recommendation feeds, and their effects are still unfolding.
What separated the decisions that worked
- Designs that removed a step users already wanted gone (News Feed, the touchscreen, the algorithmic feed) spread quickly; designs that guessed at what users wanted and interrupted them (the Office Assistant) failed.
- Optimising a design for an easily measured signal, such as watch time or a thumbs-up, repeatedly produced side effects the company later had to correct.
Where this evidence is weak
Many design decisions have no process grade, because the evidence of how they were reasoned is thin. The outcomes are better documented than the deliberations.
By who decided
| Decision-maker | Cases | Paid off | Mixed | Backfired | Too early |
|---|---|---|---|---|---|
| Market leader | 12 | 1 | 3 | 3 | 5 |
| Challenger | 3 | 2 | 1 | 0 | 0 |
| Young company | 9 | 6 | 3 | 0 | 0 |
| Industry-wide | 2 | 0 | 1 | 0 | 1 |
| Researchers and communities | 3 | 2 | 1 | 0 | 0 |
The cases
Each links to the essay, memo or road not taken where the decision is described and judged, with its sources.
- Link notes to each other instead of filing them under headings Design for connection, not just storage; the value of a collection is in what it returns that you did not expect.
- Rank pages by the links pointing to them, and offer a plain search page rather than a portal Look for signals your users already produce honestly, and protect the core task from the clutter that business pressure adds around it.
- Make the routine decisions once, for everyone (“convention over configuration”) Identify the decisions that do not deserve your users’ attention, and make them for them.
- Let anyone upload video and remove infringing material only on notice, then build a system that matched uploads to owners' works and let owners block them or take a share of the advertising Permissive growth can be made durable if a way to share the gains with the people harmed arrives before the courts do.
- Make the shared document the meeting place Moving communication into the shared object can reduce messages rather than speeding them up.
- Turn profiles into News Feed, a continuous stream of friends’ activity When a feature exposes users more, ship the controls with it, not after the protest.
- Remove the keyboard and buttons and make the whole front a touchscreen Replacing fixed hardware with software can make a device general; check which of your fixed features users would trade for flexibility.
- Rebuild the product four times around blocks that can become databases If users build their own tools with your product, you have a platform, not an app.
- Build a video app around a feed chosen entirely by algorithm from all content, rather than from accounts the user follows A product that works for a new user on the first day can outgrow one that needs the user to build something first.
- Copy Snapchat’s Stories format openly Novelty is not a moat. If your advantage is a format, assume a larger competitor will ship it within a year.
- Choose every video with a recommendation model instead of from the people you follow Ask whether the moat everyone believes in is really what users value, or only what incumbents have.
- Design an English-like query language for non-programmers Your product may succeed with a different audience than you planned. Watch who actually uses it.
- Show everyone whether you are online (presence) Every signal you expose about a user creates an expectation others will act on.
- Promise to “remember everything”: capture anything, find it anywhere Ask whether your core metric (items saved) measures the outcome users actually want.
- Give receivers control: threads, do-not-disturb, scheduled send, status messages Design for the receiver as carefully as for the sender.
- Rank and recommend videos by expected watch time rather than clicks or views Every improvement in a proxy metric creates a new way for the system to diverge from what people actually value.
- Sell “a reduction in information overload” through open, searchable channels If your product makes sending cheaper, measure the cost to receivers, not only the reduction in email.
- Re-rank the feed for “meaningful social interactions” Before you optimise a proxy, ask what the worst content that scores well on it would be.
- Fill the main feed with full-screen recommended videos from accounts users did not follow, following TikTok Changing what a product is for its existing users needs their consent, or at least their patience.
- Put an OpenAI model into its search engine within weeks of ChatGPT’s launch, then cap conversations when it misbehaved A challenger can afford public experiments that an incumbent cannot; plan the limits you will add before you need them.
- Launch with a one-hour recording time, favouring a smaller cassette and picture quality. Ask what customers will do with the product once they have it, not only what they do today.
- Add an animated Office Assistant that offered help when it guessed what you were doing Proactive features must be right far more often than passive ones to be welcome.
- Add a reward for users’ immediate thumbs-up reactions to the main ChatGPT model, then roll it back If you optimise a product for momentary approval, it will learn to flatter.
- Release a spreadsheet that separated formulas from data and organised numbers by named categories, first on NeXT A superior design on an obscure platform becomes someone else’s feature.
- Add Python and an AI assistant inside the spreadsheet When you make a powerful tool easier to use, make it easier to check as well.
- Make the word processor write: drafts, rewrites and summaries, priced into subscriptions When you automate the core activity, decide deliberately whether to do the work for users or to make them better at it.
- Add AI assistants that summarise, triage and draft messages Before automating an activity, ask whether it should happen at all.
- Let a model read a person’s notes and answer questions grounded in them When AI removes effort, ask whether the effort was the point.
- Put AI-generated answers at the top of the results page, above the links When you answer the question yourself, you take on responsibility for the answer and for the people who supplied it.
Other kinds of decision
- Opening or closing
- Business model and pricing
- Make, buy or partner
- Acquisitions and sales
- Technology bets
- New categories and markets
- Distribution and positioning
- Focus, exit and investment
- Trust, safety and privacy
- Executing large change
- Organisation and governance
- Public rules
Shrikant Kumar. “Product design.” The Telos Decision Record, Telos, version of 2026-11-03. https://shrikantk.com/record/product-design/