PTN—09
Modes
Named operating states that trade speed, depth, or autonomy — chosen once, applied to everything after.
Not every request deserves the same machine. A quick factual question and a twenty-minute research task want different depth, cost, and patience, and forcing one behavior to serve both means it serves neither. Modes let people declare the stakes up front — fast versus thorough, chat versus agent, suggest versus act — and have the whole system reconfigure to match.
A mode is a contract, which is what separates it from a tuner. A tuner adjusts one parameter of one request; a mode changes what kind of interaction you are in — how long it runs, what it may touch, how it presents results. That is why the shift must be legible: ChatGPT’s extended-thinking-style options and Cursor’s ask-versus-agent distinction both change the interface, not just a hidden flag, so people can see which contract they are under.
Modes fail through proliferation and through blur. Two or three modes with genuinely different behavior are a vocabulary; six modes with overlapping behavior are a quiz the user did not sign up for. Each mode needs a one-line promise a person can hold in their head — what it is for, what it costs, what it will not do — and the differences must be experienced in the results, not just claimed in the picker.
When to use
- Requests split into clear regimes — quick answers versus deep work — with different cost and latency.
- The AI can act with different autonomy levels, and the choice should be deliberate, not per-message.
- One interface serves both novices who want safety and experts who want power.
- People keep prefixing prompts with meta-instructions like “think hard about this” or “just do it”.
Design considerations
- 01Cap the roster at two or three modes whose behaviors are distinguishable in the output.
- 02Name modes by what they do for the person — “Research”, “Agent” — never by internal model names alone.
- 03State each mode’s cost honestly at the picker: time, credits, or level of autonomy granted.
- 04Show the active mode persistently during the interaction, not only at selection time.
- 05Pick a sensible default and let escalation be one click when a request outgrows the current mode.
- 06Change the interface with the mode — a long-running agent mode should look different from chat.
- 07Let a mode switch mid-conversation apply forward cleanly instead of forcing a restart.
Pitfalls
- ✕Mode sprawl — five overlapping options that turn every message into a routing decision the user is not equipped to make.
- ✕Indistinguishable modes: the “deep” option takes longer and bills more but produces answers users cannot tell from the fast one.
- ✕Invisible mode state — someone left agent mode on, and the next casual request starts editing files.
- ✕Marketing names in the picker (internal model codenames, vague superlatives) that tell people nothing about which to choose.
In the wild
- ChatGPT
- Deep research is a distinct mode with its own contract — it clarifies scope, runs for minutes, and returns a cited report instead of a chat reply.
- Cursor
- Ask and agent modes draw the autonomy line explicitly: one answers questions about code, the other edits files and runs commands.
- Claude
- Extended thinking is an opt-in regime that trades latency for reasoning depth, surfaced as a visible choice rather than a hidden flag.
- Perplexity
- Standard search and deeper research modes split quick lookups from long multi-source investigations, with the cost difference stated up front.