Prompting frameworks

Use the method the problem needs

Choose a framework that addresses the failure you need to fix.

Begin with the missing capability

If the model lacks facts, provide evidence. If the format is unclear, show an example or define a schema. If the work has dependent stages, decompose it. Verification is useful only when a check can meaningfully challenge the draft.

NeedMethod to consider
Relevant factsSource grounding or retrieval-augmented generation
A consistent formatFew-shot examples and structured output
Dependent stagesTask decomposition or prompt chaining
Checkable factual claimsChain of Verification
Several reasoning candidatesSelf-consistency, with added sampling cost
Authorized external actionsA tool-using workflow with explicit permissions

Choose a method for a supplier brief

Apply the parts that address the missing facts, format or checks in this brief.

Illustrative framework selection
For a supplier comparison, first collect the approved proposals. Use the same comparison table for each supplier, then identify missing values. Verify the final factual claims against the proposals. Do not generate multiple speculative rankings to compensate for missing prices. Ask for the missing commercial information before making a recommendation.

More steps can cost more without helping

Additional samples, tool calls and review passes can increase latency and cost. Test the chosen method on representative cases. Research results for a particular model and dataset do not guarantee the same improvement in your workflow.

Common questions

Should I always request a detailed chain of thought?

No. Ask for the explanation, assumptions, evidence or calculation needed to check the result. A long reasoning narrative is not itself proof of correctness.

Sources and review

Reviewed

Put the workflow to work.