Lesson 20 of 25
The Strategic Use of Temperature
Temperature is the setting that controls how adventurous a model's word choices are: low makes it predictable and literal, high makes it surprising and inventive. The important thing to know is that you cannot touch it in the ChatGPT, Claude or Gemini apps - there is no slider. It is an API setting, or one you adjust in Google AI Studio. In the apps you get the same effect through wording: "stick strictly to the text, add nothing" pulls the output one way, "be surprising, give me unexpected angles" pulls it the other.
This is the one platform where the setting is genuinely reachable - not in the Gemini app, but in Google AI Studio, where a temperature slider sits beside the prompt. That makes it the right place to actually see the difference: run the same prompt at a low value and a high one and read them side by side. In the Gemini app itself there is no slider, so you steer with wording exactly as elsewhere, and a saved Gem can hold a standing "literal, source-bounded" instruction for work where that always applies.
- The slider is in Google AI Studio, not the Gemini app.
- Compare the same prompt at low and high there once.
- For a strict summary, say "do not search" explicitly.
- Keep a "literal, source-bounded" instruction in a Gem.
- High temperature invents facts, not only better phrasing.
Perplexity version
### GOAL ### A precise, factual summary of a news article. ### A NOTE ON TEMPERATURE ### There is no temperature setting in the Gemini app. If you want to actually set it, open Google AI Studio and use the slider there - this prompt is worth running at a low value and again at a high one just to see what the difference looks like on the same text. In the app, the instructions below do that job instead. ### TASK ### Summarise the five key facts of the attached article as a numbered list. I am attaching the file rather than pasting it - work only from that. ### DO NOT SEARCH - THIS IS THE CRITICAL INSTRUCTION ### Do not ground this in Google Search. Not for context, not to check a figure, not to see how other outlets reported it. Grounding is the specific way a factual summary goes wrong on this platform: material arrives from outside my source and reads as though it came from my document. Everything in your answer must be traceable to a line in the attached file. ### CONSTRAINTS ### * Only what is in the text. No background, no causes the article does not state, nothing you know that the article omits. * No interpretation - not what it means, who it favours, or what follows. * Plain language. "The minister said X", not "the minister insisted X". * Names, numbers, dates and quotes exactly as written. No rounding, no converting lakh, crore or million. * Attribute as the article attributes. "Officials said" stays "officials said". * Fewer than five distinct facts? Give fewer and say so. ### AFTER THE LIST ### 1. "What I left out" - anything you judged opinion, analysis or framing rather than fact. 2. "Where I had to choose" - any sentence whose plain meaning required a judgement call. Name it. ### OUTPUT ### Into a Google Doc called "Article summary", each fact followed by the paragraph it came from, so every line is checkable. Save these constraints as a Gem called "Strict factual summary" - I do this often and the no-search rule is the part I would forget to retype.
Perplexity version
### GOAL ### Brainstorm creative names for a new food delivery service in Hyderabad. ### A NOTE ON TEMPERATURE ### This is the prompt to try in Google AI Studio with the slider pushed high, because the difference is visible immediately on a naming task - the same prompt at a low value returns the obvious names and at a high one returns the strange ones. The Gemini app has no slider, so the wording below asks for it instead. ### THE BRIEF ### Act as a creative branding expert. A food delivery app in Hyderabad whose single promise is speed - faster than anything else in the city. ### DO NOT SEARCH ### No Google Search for the generation. Searching food delivery brand names returns the names that already exist and rank, which is the exact opposite of what a brainstorm is for. ### THE TASK ### 20 unconventional, catchy names. Be genuinely adventurous - play with words, blend English with Telugu, Hyderabadi Urdu or Hindi, invent compounds, use sound. I would rather reject five as too strange than forget twenty. Spread them across five approaches, four each - this is what stops "be creative" producing twenty synonyms for fast: 1. Speed by implication - fast without the word. 2. Telugu, Urdu or Hindi words used straight, with meaning and pronunciation. 3. Invented words - coinages, portmanteaus, sounds that feel right and mean nothing. 4. Deliberately wrong-footed - nothing obviously to do with food or speed. 5. Hyderabad-specific - something only this city would name a company. **Banned:** fast, quick, rapid, swift, zoom, dash, flash, speed, express, instant, blitz, turbo, rocket, now, bite, eats, food. ### THEN SEARCH - FOR THE CHECKING ONLY ### For your top eight, now use Google Search and tell me with citations: whether the name or something close is already used by a food, delivery or logistics brand in India; whether the .in and .com domains look taken; whether it collides with a film title, a political slogan or an existing FMCG brand. For the Telugu and Urdu names, confirm the meaning against a dictionary source rather than recollection, and say if you cannot. Be explicit about what search could not settle - trademark registration is not reliably searchable this way and I need to know that rather than assume you checked. ### OUTPUT ### A Google Sheet called "Delivery app names" with columns: Name, Approach, Meaning, Pronunciation, Search findings, My score. Leave the last column empty - I will sort on it.
Perplexity version
### GOAL ### Both a factual summary and a creative interpretation of a poem. ### A NOTE ON TEMPERATURE - READ THIS FIRST ### The original version of this exercise implies you can turn a dial down for the first task and up for the second. You cannot do that in the Gemini app, which has no temperature control at all, and you cannot change it partway through a conversation anywhere. In Google AI Studio you could run the two halves as two separate prompts at two different settings, and doing that once is genuinely instructive. Here, the two tasks are separated by their instructions instead. ### SUBJECT ### "The Tyger" by William Blake. ### TASK 1 - LITERAL ### One clear paragraph on what the poem says at the surface: who speaks, what they look at, what they ask. Quote its own words where that helps. * No symbolism, no "represents", no "stands for". * Nothing about Blake's life, *Songs of Experience*, or "The Lamb". * **Do not search for this task.** Search "The Tyger" and you get a century of interpretation, which is precisely what Task 1 excludes. Work from the poem itself. ### TASK 2 - INTERPRETIVE ### Three readings of what the Tyger might signify. Be imaginative and go somewhere unexpected. * The three must be genuinely incompatible, not three angles on one reading. Three versions of "divine power" is a failed answer. * Range widely - theological, political, psychological, aesthetic, or something I have not considered. The industrial revolution and the problem of evil are the standard two; at least one of yours should be strange. * Anchor each in specific lines however far you take it. * For each, name what it cannot explain. Which lines resist a reading is the interesting part. ### FOR TASK 2 ONLY, SEARCH ### After you have given me your three readings, use Google Search once: which of them are already established in the critical literature, and which are genuinely yours? Cite what you find. I would like to know whether what I am reading is a known reading or a new one - and if one of your three turns out to be a well-worn critical position, say so plainly rather than letting it look original. ### THEN ### One short note: where the boundary between the two tasks was hard to hold. "Fearful symmetry" is not a neutral description of a tiger, so a strictly literal summary of this poem may not be possible. If Task 1 forced you to interpret in order to describe, name the line that did it. ### OUTPUT ### Keep it in the chat - I will work through it rather than file it.