Searchable prompt library

Prompts that ship.

Hand-picked, model-tagged prompts you can copy and use right now. Search, filter, copy — or ship one as an app in a click.

You are a router. Classify the task below as TRIVIAL or HARD.
TRIVIAL = answerable in one pass with no tools, no math, no ambiguity.
HARD = anything else.

Compare ANSWER against SOURCE. For each factual claim in ANSWER,
mark SUPPORTED (verbatim in SOURCE), INFERRED (reasonable), or UNSUPPORTED.
Return JSON: {"verdict":"pass|fail","unsupported":["..."]}.

Your previous reply did not parse as JSON. The parser said:
<<<error>>>
Return ONLY the corrected JSON. It must satisfy this schema:

Wrap the prompt below into a single-file app spec.
Output: (1) input fields with types, (2) the system prompt with {{slots}},
(3) the output schema, (4) one happy-path example.

Summarize this diff for a reviewer. Output exactly:
- **What changed** (3 bullets max, behavior not files)
- **Why** (1 line)

Write a cold email. Constraints:
- Under 90 words. Subject under 6 words.
- First line references something specific about the recipient (use CONTEXT).

Extract fields from TEXT into this schema. Use null when a field is absent.
Never guess. Never invent. Copy values verbatim where possible.
Return JSON only, matching the schema key order exactly.

Score the SUBMISSION against each rubric criterion from 0-5.
For each: {criterion, score, reason (<=15 words)}. Then a total.
Anchor scores to the rubric text, not vibes. Same input → same score.

Given the QUESTION, produce 5 web search queries that, combined, cover it.
Each query attacks a different angle (definition, counter-evidence, recency,
primary source, expert critique). No two should return the same results.

Try to REFUTE the claim below using only the provided SOURCES.
List the strongest disconfirming evidence first. Then rule:
SUPPORTED / CONTRADICTED / INSUFFICIENT. Default to INSUFFICIENT.

Before calling any tool, output a PLAN: the ordered tool calls you intend,
each with its purpose and the exact arguments. Stop after the plan.
Do not call tools yet. If no tool is needed, say 'NO TOOLS' and answer directly.

Rewrite the TEXT to be ~40% shorter with identical meaning.
Kill hedges (just, really, very, I think), filler, and repeated ideas.
Keep the author's voice and every concrete fact. Return only the rewrite.

Write ONE read-only SQL query answering the QUESTION against SCHEMA.
SELECT only — never write, never DROP. Use explicit column names.
If the question can't be answered from SCHEMA, return: -- cannot answer

Turn these commits into a changelog for end users.
Group under: Added, Changed, Fixed. Drop chores, merges, and refactors.
One line each, present tense, benefit-first. No commit hashes.

Summarize DOC for a decision-maker in under 120 words:
1. The one takeaway (1 sentence).
2. Three facts that matter, most load-bearing first.

Generate 4 few-shot examples for the TASK below.
Cover: a typical case, an edge case, an ambiguous case, and a failure case.
Each example: {input, ideal_output}. Outputs must follow the task's format exactly.