The manifesto
What it takes for a prompt to be trusted with real work.
Every prompt in this library follows seven principles. They exist because a prompt's output often ships without anyone checking it first — so the prompt has to be its own skeptic, its own QA gate, and its own compliance flag.
Contracts everywhere
Every prompt declares inputs and a machine-parseable output, so any prompt's output can feed another prompt. That's what makes 107 prompts an operating system instead of a pile.
Honesty valves
Prompts are allowed to return INSUFFICIENT_DATA, REJECT, or "don't run this campaign" instead of forcing an output. Nobody reviews every output before it ships, so the prompt has to be its own skeptic.
Evidence binding
No invented stats, quotes, cultural claims, or persona traits; everything cites its input or gets flagged [NEEDS SOURCE] / [NEEDS NATIVE REVIEW] / [LEGAL REVIEW].
Industry-agnostic variables
Audience, offer, economics, and jurisdiction are parameters, so the same prompt serves SaaS, CPG, local services, or ecommerce.
Escalation gates
Compliance, legal, customer-approval, and human-review flags are built into outputs, defining exactly where the human belongs in the loop.
Kill criteria
Budget, experiment, and channel prompts pre-commit to the number that stops the activity. Machines shouldn't have sunk-cost bias, and these prompts don't.
Model routing
Haiku for gates/classification/governance, Sonnet for judgment/writing, declared per prompt with cost in mind.
See the principles applied — every prompt page shows its contract, its failure modes, and its version history.
Browse the library