Constraining output structure
The single highest-leverage advanced technique is forcing the model to produce structured output. Instead of asking for a “summary”, specify an exact JSON schema — field names, types, and allowed values — and instruct the model to emit only valid JSON. Modern APIs from OpenAI and Anthropic support a JSON or tool-calling mode that guarantees parseable output. Structured output makes responses machine-readable, eliminates ambiguity, and turns the model into a component you can wire into a pipeline rather than free text you must re-parse.
Role assignment and meta-prompting
Assigning a precise role (“You are a senior tax accountant reviewing a UK limited-company return”) primes the model toward the right vocabulary, depth, and caution. Push this further with meta-prompting: ask the model to first write the ideal prompt for a task, critique it, and revise it before executing. This is especially effective when you do not yet know the best framing — you let the model surface the constraints and edge cases you would otherwise miss, then run the refined prompt.
Self-consistency and self-critique
For problems with a checkable answer, self-consistency beats a single pass. Sample several independent chain-of-thought reasoning paths at a moderate temperature, then take the majority vote on the final answer. Different paths cancel out individual mistakes. A complementary pattern is the self-critique loop: have the model produce a draft, then in a second turn explicitly ask it to find flaws in its own answer and fix them. Naming the failure modes to look for (“check the arithmetic, verify each citation exists”) makes the critique far more effective than a vague “improve this”.
Program-aided language models
LLMs reason well but compute badly. The program-aided language model (PAL) pattern fixes this by having the model write code to handle the exact calculation, which is then executed externally. Ask for the answer “by writing and running Python”, and the model offloads arithmetic, date math, and data transformations to a deterministic interpreter. This is the foundation of code-interpreter tools and removes a whole class of confident-but-wrong numerical answers.
Composing techniques and knowing when to stop
These techniques compose: a production prompt might assign a role, demand a JSON schema, use chain-of-thought internally, and run a self-critique pass. But each layer adds cost and latency. The discipline is to add complexity only where a simpler prompt measurably fails. Start with a clear instruction and a few examples; reach for self-consistency, PAL, or critique loops only when you can see the errors they are meant to fix. Measure against a small evaluation set so you know each addition is actually helping rather than just feeling sophisticated.