Copy-and-paste · verified July 27, 2026
Research and analysis prompts for Claude Fable 5
Six scaffolds for synthesis, comparison and evidence work, built around telling the model what it may treat as known.
What actually helps here
State the source boundary before the question. A model asked to analyse will analyse, and if the material is thin it fills the gap from training data without flagging it. Say which documents count, ask for a confidence label on each claim, and require that anything outside the supplied sources be marked as such.
The prompts
Fill in the bracketed slots, paste into any Claude app or the API. Written for Fable 5's goal-first style — they state the outcome and the constraints rather than scripting the steps.
Structured comparison
Compare [option A] and [option B] for [use case]. Return a table with criteria rows: cost, learning curve, ecosystem, risks. End with a 2-sentence verdict and a confidence level.
Assumption audit
List every assumption hidden in this claim: "[claim]". Mark each as testable or untestable, and suggest how to verify the top three.
Balanced summary
Summarize the key arguments for and against [topic] in under 250 words. Attribute each argument to its typical proponents and note where the evidence is weakest.
Pros, cons, verdict
Give pros, cons, and a final verdict on [decision]. Limit to 5 bullets each, ranked by importance to [stakeholder].
Data interpretation
Given these figures: [paste data], describe three trends, one anomaly, and one question the data cannot answer. Avoid speculation beyond the numbers.
Steelman both sides
Steelman the strongest case FOR and AGAINST [position] in two 150-word passages, then identify the crux that actually separates them.
Getting more out of these
Two habits matter more than the wording. Replace every bracketed slot before you send — Fable 5 follows instructions literally, and an unfilled [topic] becomes part of the task. And resist adding "think step by step" or "show your reasoning": on Fable 5 that phrasing can trip the reasoning_extraction classifier and hand your request to Opus 4.8 instead.
Need something these do not cover?
The generator builds a full system prompt around your own task, autonomy level and output style, and hands back the API call to go with it.