Verified against GitHub Copilot Chat · 2026-08-06
Choose the right model in Copilot Chat's model picker instead of leaving it on default for everything
A task-classification prompt that states the actual reasoning demand of the work before picking a model from Copilot Chat's picker, so a quick mechanical rename and a genuine architectural tradeoff are not both run on whichever model happened to be selected last.
The prompt
Ready to copy — highlighted parts are example details you can swap.
Before sending this to Copilot Chat, decide which model in the model picker this task actually needs — do not leave it on whatever was selected for the previous, possibly very different, task. TASK Decide whether to store saved searches as a new database table or as a JSON column on the existing user_preferences row, and justify the choice. REASONING DEMAND Genuine tradeoff reasoning — this is an architectural decision with real long-term cost either way, not a mechanical task. MODELS AVAILABLE IN THIS WORKSPACE'S PICKER Claude Sonnet 4.6, GPT-5.2, and Gemini 3 Pro are all enabled in this workspace's picker. LATENCY OR COST CONSTRAINT None for this task — it runs once, so a slower, stronger model is worth the extra wait. PICK USING THIS RULE If Genuine tradeoff reasoning — this is an architectural decision with real long-term cost either way, not a mechanical task. describes genuine tradeoff reasoning — comparing two real architectural approaches, tracing a subtle bug across several files, reviewing a security-sensitive change — select the strongest reasoning-oriented model Claude Sonnet 4.6, GPT-5.2, and Gemini 3 Pro are all enabled in this workspace's picker. offers, even if it is slower, because the cost of a wrong architectural call caught late is categorically higher than the extra seconds a stronger model takes to answer. If the task is mechanical — a rename across files matching an exact pattern, formatting a list into a table, generating a boilerplate test file from an existing template — a faster, cheaper model in the picker will produce an equivalent result with less latency, and running it on the strongest available model for that kind of task spends budget the task did not need spent. BEFORE SWITCHING MODELS MID-CONVERSATION State explicitly that the model is changing and why, in the same message where you switch — a different model does not share the exact reasoning trace of the model that answered the previous turn, only the visible conversation text, so a silent mid-conversation switch can produce an answer that looks like it continues the previous reasoning but was actually regenerated from scratch by a different model reading only the transcript, not the reasoning that produced it. If the strongest model is rate-limited, fall back to the second-strongest reasoning model rather than the fastest one, and say explicitly that a fallback happened. CONSTRAINTS - Do not assume every model in Claude Sonnet 4.6, GPT-5.2, and Gemini 3 Pro are all enabled in this workspace's picker. handles the same prompt structure equally well — a reasoning-oriented model generally tolerates an underspecified, exploratory question better than a faster completion-oriented model, which tends to need the task, constraints, and output format spelled out explicitly to produce a comparably good answer; adjust how much structure you give the prompt to the model actually selected, not a generic phrasing meant to work identically on all of them. - If None for this task — it runs once, so a slower, stronger model is worth the extra wait. rules out the model Genuine tradeoff reasoning — this is an architectural decision with real long-term cost either way, not a mechanical task. would otherwise call for, say so explicitly rather than silently using a weaker model and treating its answer with the same confidence you would give the stronger one — a constrained choice should come with a correspondingly more skeptical read of the output, not the same trust level.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
Different models genuinely differ in reasoning depth versus latency and cost, which makes the model picker a real, consequential per-task choice rather than a cosmetic setting — the actual tradeoff this prompt asks you to weigh explicitly is that a wrong call on a genuine architectural decision, caught late after code has already been built on top of it, costs categorically more than the extra seconds a stronger model takes to answer, while a mechanical task run on the strongest available model spends latency and cost budget on work that did not need that level of reasoning at all. A mid-conversation model switch does not carry over the reasoning trace of the model that answered the previous turn — only the visible conversation text does — which means a silently swapped-in model reads the transcript as its entire memory of the conversation and reconstructs its own reasoning from scratch, producing an answer that can look like a natural continuation while actually being a fresh, independent judgment that never saw whatever led to the earlier turn's specific phrasing or caveats; stating the switch explicitly at least surfaces that discontinuity instead of letting it pass unnoticed. Prompt structure needs genuinely vary by model type in a well-documented way: a reasoning-oriented model tends to tolerate an underspecified, exploratory question and do useful work filling in the gaps itself, while a faster, completion-oriented model tends to need the task, constraints, and output format spelled out explicitly to produce a comparably reliable answer, so a prompt written with one model's tolerances in mind can underperform when the same wording is sent to the other. Finally, calibrating trust to the constraint that actually determined the choice — treating an answer from a cost-constrained fallback model with more skepticism than one from the strongest available model — matters because the alternative is applying uniform confidence to outputs that were not produced under uniform conditions, which quietly erases the very information a stated fallback was supposed to preserve.
Verified against
GitHub Copilot Chat Copilot Chat 1.271 (VS Code, multi-model picker) · 2026-08-06
Changelog
- 2026-08-06 — Initial publish, verified against Copilot Chat 1.271 (VS Code) multi-model picker.
Need this built into your business?
If a prompt isn't enough — custom software, built and maintained for you — that's Scult's day job.
EXPLORE CUSTOM SOFTWARE
