Every prompt recipe. Copy, paste, ship.
48 posts
OpenAI prompt caching is implicit, so the failure mode is not a cache you never made, it is one you keep destroying. Four recipes, the request settings that void a prefix without touching your text, and the two token floors the vendors disagree on.
Maieutic prompting makes a model argue both sides, then keeps only the explanations that survive. What the 2022 paper really does, the scope it covers, and 4 copy-paste recipes.
An LLM confidence score tells you how sure the model was. Which providers return logprobs in 2026, what to do on Claude which returns none, and 4 copy-paste recipes.
Five copy-paste Tree of Thoughts (ToT) prompt recipes for 2026: the one-prompt expert panel, branch-score-pick, explicit backtracking, ToT for planning, and when to skip ToT and just use self-consistency.
Five copy-paste top_p (nucleus sampling) recipes for the OpenAI and Anthropic APIs in 2026: the default that ships, clamping the tail for extraction, focused creative, steadier JSON, and the one-knob rule.
Four copy-paste logit_bias recipes for the OpenAI API in 2026: ban a token, pin a classifier, fix the tokenizer mismatch, and know when it is the wrong tool.
Count Claude tokens before you send with the free count_tokens endpoint: pre-flight the context window, estimate a batch bill, and recount after a model swap.
Prompt templates with {{variables}} in 2026: 5 copy-paste recipes for reusable, cacheable, injection-safe prompts on Claude and OpenAI, plus the failure modes that bite.
max_tokens caps output; it does not push the model to fill it. Detect truncation via stop_reason of max_tokens, continue a cut-off answer, and leave room for extended thinking. Five copy-paste recipes for 2026.
Streamed tool arguments arrive as partial_json fragments, not JSON. Four copy-paste recipes to buffer by index, parse once, and stop corrupting tool calls, on Claude and OpenAI.
The Message Batches API runs LLM jobs async for 50% off in Claude and OpenAI. Four copy-paste recipes: submit, poll without hammering, match results by custom_id, and survive the 24-hour expiry.
frequency_penalty counts, presence_penalty flags. Four copy-paste recipes for when to grab which OpenAI sampling knob, plus the one config that quietly breaks your JSON.
