The Universal Prompt Contract
There is no perfect template for all models - there is a set of engineering questions the contract answers. A section is worth including only if it changes behavior or makes the result verifiable; everything else is noise.
Here are the questions any prompt is assembled from.
- Goal - what user result is needed?
- Boundaries - what does the system do and, above all, not do?
- Data roles - which fields are trusted, which are user-supplied?
- Context - what facts may be used?
- Decision rules - what heuristics and priorities are needed?
- Tools - when to call, when to clarify, when not to?
- Output - text, schema, citations, language, and length?
- Done - what counts as success, what as blocked, what as escalate, and what is the effort limit?
Put together, they give a skeleton that repeats from task to task.
# Goal
{{what the user should get}}
# Boundaries
- Do: {{the allowed area}}
- Don't: {{what belongs to another process}}
# Context and sources of truth
{{data, priorities, freshness, trust}}
# Decision rules
{{only rules that can't be reliably derived from the context}}
# Tools
{{when to call, required parameters, confirmation}}
# Result
{{schema or a clear format description}}
# Uncertainty and completion
{{when to clarify, refuse, escalate, stop}}To feel how the sections fit a specific scenario, assemble the contract in the builder.
Interactive lab 2
Prompt contract builder
# Goal Give a grounded answer from the passed sources # Scope Describe what is allowed and explicitly excluded. # Context retrieved sources with source_id and effective_date # Decision rules Add only rules that fix observed errors. # Output Strict schema + domain validation # Risk control Timeout, validation, and an ordinary fallback # Done answer, clarify, not_found, or conflict
And a word on role separately - it is useful exactly as far as it carries a contract.
A role is useful if it carries a contract. "You are an editor of medical translations who preserves the glossary terminology and adds no diagnoses" sets behavior. "You are a brilliant expert" defines almost nothing.