The Prompt Lab — Granularity Targeting

Granularity Targeting

The Technique

Granularity Targeting means explicitly specifying the level of detail you want in a response — not just the topic, but how deep, how broad, and at what resolution. Most models default to a middle-ground level of detail that satisfies no one: too shallow for experts, too dense for beginners, and almost never calibrated to how the output will actually be used.

The Naive Prompt

Explain how Claude Fable 5.1's adaptive thinking works.

Why It Falls Short

Without a granularity signal, the model produces a generic encyclopedia entry — covering everything lightly but committing to nothing deeply. You get a paragraph that would bore an AI engineer and confuse a product manager equally. The prompt gives zero information about whether you need a two-sentence briefing, a technical deep-dive, or something in between.

The Improved Prompt

Explain how Claude Fable 5.1's adaptive thinking works. Target this at a 
technical product manager who understands transformer basics but hasn't 
read the model card. I need exactly three things: (1) what "always-on 
adaptive thinking" means mechanically in one crisp sentence, (2) how it 
differs from the optional thinking modes in Gemini 3.8 Flash, and (3) one 
concrete example of a task where this changes the output quality. 
No background history. No pricing. Stop after the example.

Why It Works

The improved prompt specifies audience expertise, enumerated scope (three items, no more), explicit exclusions (no history, no pricing), and a hard stop — all of which are granularity levers. The model now knows it shouldn’t build up to the answer or wind down after it; the output density is set from the first token. Enumerated scope is especially powerful because it forces the model to allocate its “effort budget” across exactly the items you care about rather than spreading it across whatever it judges to be relevant.

When to Use This

• When you’re on a deadline and scanning outputs — tight granularity means you read the response once, not three times looking for the useful sentence buried in the middle.
• When cost scales with output length — on models like GPT-6 Astra ($50/M output) or Claude Fable 5.1 ($50/M output), an unfocused prompt that generates 1,200 words when you needed 200 is a literal billing problem, not just an aesthetic one.
• When you’re chaining prompts in an agentic pipeline — downstream prompts often fail because upstream outputs contain the right information at the wrong granularity; a too-verbose summary confuses the next step just as much as a too-sparse one.