Prompt Engineering

Source: dair-ai/Prompt-Engineering-Guide at commit 5767372 (2026-03-11), English pages, notebooks, and lecture slides. Everything below was read in full and audited. Reference files carry templates, examples, caveats, and per-file "Audit notes".

Read first: the guide is mostly a 2022 to early-2024 snapshot (GPT-3.5, text-davinci-003, GPT-4, Mixtral era). Technique concepts hold up. Model-specific claims, prices, benchmarks, and code samples do not. See references/audit-summary.md before trusting any date-bound claim.

Workflow for writing or fixing a prompt

1. Define the task, the output format, and how you will judge success. Without a check, you can't tell if a rewrite helped.
2. Start with the simplest prompt that could work: clear instruction, specific context, delimiters, explicit output format. Say what to do rather than what to avoid. Specificity beats length.
3. If output is wrong, escalate one step at a time: add few-shot examples (format and label space matter more than label correctness) -> reasoning (CoT, or a reasoning model) -> decomposition (prompt chaining) -> tools or retrieval (PAL, ReAct, RAG) -> fine-tuning.
4. Set parameters deliberately: low temperature/top-p for factual or extraction tasks, higher for creative tasks. Change one of temperature or top-p, not both.
5. Test against adversarial and edge inputs, including prompt injection if the prompt handles untrusted text.
6. Iterate on a fixed test set. A single lucky run proves nothing. The repo's own bias and exemplar-order examples are single-run, so don't copy that.

Where to look

Load only the file the task needs. Each starts with a table of contents.

Quick technique picker

• Simple known task -> zero-shot with an explicit format.
• Format or labels wrong -> few-shot.
• Multi-step reasoning on a non-reasoning model -> CoT ("Let's think step by step" or worked exemplars); unstable answers -> self-consistency (sample N, majority vote).
• Exact arithmetic or dates -> PAL (generate code, execute in a sandbox).
• Fresh, private, or hallucination-prone facts -> RAG, or ReAct with search.
• Too complex for one prompt -> prompt chaining.
• Search or planning puzzles -> Tree of Thoughts.
• Agent repeats mistakes with a feedback signal -> Reflexion.
• Reasoning models (o-series style): give a clear goal and constraints and skip manual CoT. The repo says so, though other pages recommend CoT generally. Treat CoT as for non-reasoning models.

Cautions

• Treat repo example outputs as illustrations, not benchmarks. Many came from retired models at temperature 1.
• Known bugs in repo examples: the PAL date exemplar returns today.strftime instead of later.strftime; the self-consistency bagels exemplar lacks its "The answer is" line; the function-calling notebook double-encodes tool arguments. The reference files carry corrected notes.
• Never exec() model output outside a sandbox (the PAL notebook does).
• Several technique pages (DSP, GraphPrompts, Active-Prompt, ART, multimodal CoT, meta-prompting) are stubs. The reference sketches for them are tagged "(not in repo)".
• Gaps in the guide: indirect prompt injection via tool output, MCP, structured outputs, evals, least-to-most, self-refine, chain-of-verification, chunking and reranking for RAG. Use current provider documentation for these.
• Agent and context-engineering pages are one course's opinion and carry promo callouts. Their prompt examples contradict each other (subtask counts, schema types).