AI Prompt Templates That Work

Summarize document

Tuned for GPT-4o: a named reader, a hard length cap, and an extractive-only rule. GPT-4o is concise and obedient to a word limit, so the prompt spends its instructions on what NOT to do — no interpretation, no ‘this document discusses’ opener.

Write email

Tuned for GPT-4o: facts and the ask as bullets, the relationship named, output capped to subject plus one screen in a stated register. GPT-4o matches a specified tone well and will cut the filler if the constraint is explicit.

Debug code

Tuned for GPT-4o: a short system-style spec with fenced delimiters and a fixed root-cause / minimal-diff / risks output. GPT-4o follows a terse numbered contract closely and degrades on long, chatty prompts, so this one bans preamble and generic advice outright.

Write SQL query

Tuned for GPT-4o: schema as DDL in a fenced block, the grain stated plainly, output limited to the query plus one-line assumptions. GPT-4o writes correct SQL fast when the target dialect and join keys are pinned and it is told not to explain.

Creative writing

Tuned for GPT-4o: a crisp brief — form, POV, tone, length, one hard constraint — and an instruction to open mid-scene. GPT-4o writes cleanly to a tight spec and tends to add a warm-up paragraph unless told plainly not to.

Summarize document

Tuned for Claude 3.5 Sonnet: a <document> tag, a key-claims pass with section references first, then the summary, then an <uncertain> list of anything the source only implies. Claude is good at that fidelity check and will actually flag its own overreach when asked.

Write email

Tuned for Claude 3.5 Sonnet: <context>, <goal>, <must_include> and <tone> tags, plus a callout of the one sentence most likely to be misread. Claude is strong at diplomatic phrasing and the misread-check catches the line that would land wrong.

Debug code

Tuned for Claude 3.5 Sonnet: XML tags for the code and failure, a forced <scratchpad> execution trace before any answer, and an explicit no-refactor rule. Claude honours structured tags and negative constraints unusually well, and the scratchpad is where it actually finds the root cause instead of guessing.

Write SQL query

Tuned for Claude 3.5 Sonnet: <schema> and <question> tags plus a forced <thinking> pass over join cardinality, result grain and NULL handling before the query. That reasoning step is where Claude catches the fan-out and double-count bugs a one-shot prompt misses.

Creative writing

Tuned for Claude 3.5 Sonnet: a rich <style> block for voice and rhythm, an <avoid> list of specific clichés, and an instruction to commit to one narrator. Voice control is Claude’s strength, and naming the exact phrases to avoid works better than ‘don’t be generic’.

Summarize document

Tuned for Gemini 1.5 Pro: paste the whole document however long, get a fixed template — TL;DR, decisions, open questions, and figures quoted verbatim. The long context handles the full text; the verbatim-numbers rule stops the paraphrase drift that garbles amounts and dates.

Write email

Tuned for Gemini 1.5 Pro: structured inputs and two variants — direct and warm — each with its own subject line, under a stated word cap. Two register options plus a pick-one guide turns Gemini’s verbosity into a useful choice.

Debug code

Tuned for Gemini 1.5 Pro: paste whole files, not a snippet, because the long context window can hold them. A strict template with a single ranked cause list plays to Gemini being thorough while the explicit brevity cap stops it padding the answer.

Write SQL query

Tuned for Gemini 1.5 Pro: paste the entire schema dump, get one query plus up to two ranked alternative shapes. The big context lets it reason over every table, and the ranked-alternatives format channels its breadth into something useful instead of a wall of text.

Creative writing

Tuned for Gemini 1.5 Pro: a structured spec, two labelled openings to choose between, and an explicit ban on stacked adjectives and simile pile-ups. Gemini leans ornate, so the constraint is doing real work here.