The Gaiish Method

1. IntentWhat are you trying to accomplish?
2. ContextWhat does the AI need to know?
3. InstructionWhat should the AI do?
4. ConstraintsWhat rules, limits, tone, format, requirements or exclusions apply?
5. ResultWhat should the finished output look like?
6. ValidationHow should the AI verify the result satisfies the request?

Intent

What are you trying to accomplish?

Intent is the outcome behind the request, not the request itself. "Summarise this report" is an instruction; "decide whether to renew this contract" is an intent. State it and the model can judge which details matter — omit it and every detail looks equally important.

• Name the decision, deliverable or change you are working toward.
• Say who the output is for and what they will do with it.
• One intent per prompt. Two intents produce a shallow attempt at both.

Context

What does the AI need to know?

Context is everything the model cannot infer: the documents, the situation, the audience, the decisions already made. A model with your context outperforms a larger model without it.

• Paste the material rather than describing it.
• Include the constraints of the situation, not just the task — team size, deadline, systems in use.
• Say what has already been tried and rejected, so the model does not propose it.
• When the corpus is too large to paste, retrieval puts the right passage in the prompt for you — see embeddings and RAG.

Instruction

What should the AI do?

The instruction is the verb: analyse, draft, rewrite, extract, classify, compare. Keep it single and explicit, and number the steps when the work has an order.

• Use one primary verb; if you need several, number them.
• Say what to do, not only what to avoid — "be concise" is not an instruction, "three sentences" is.
• Separate the instruction from the material so the model cannot confuse the two.

Constraints

What rules, limits, tone, format, requirements or exclusions apply?

Constraints are the boundary of an acceptable answer. They are also where most disappointment originates, because an unstated constraint is invisible until it is violated.

• Quantify: word counts, row counts, time periods, currency, reading level.
• State exclusions explicitly — what to leave out is an instruction too.
• Name the tone and register rather than adjectives like "professional".
• Say what to do with uncertainty: report gaps rather than fill them.

Result

What should the finished output look like?

Describe the artefact, not the topic. A named shape — sections, columns, a schema — is the difference between an essay and something you can use directly.

• List the sections or columns you expect, in order.
• For machine-read output, give the exact schema and forbid commentary.
• Paste an exemplar when you have one: one good example beats a paragraph of description.

Validation

How should the AI verify the result satisfies the request?

Validation asks the model to check its own output against the request before presenting it, and to be explicit about what it could not do. It is the component most often left out entirely.

• Ask it to check the output against each constraint and report violations.
• Ask it to list the assumptions it made and the facts it could not find.
• Require citations to the supplied material for factual claims.
• Remember that self-checking is not proof — verify anything irreversible yourself.

The method in one prompt

Related