@weaveintel/prompts

Versioned prompt templates with safe variable substitution, reusable fragments, output contracts, and A/B experiments.

Why it exists

A prompt is really a document that many people edit and ship to production — and like any shipped document, an unreviewed change can quietly break things downstream. Think of it like a recipe card kept in a binder: each card has a version, blanks you fill in ("2 cups of {{flour}}"), and shared notes you paste in from other cards. prompts gives you that binder — templates you can fill in without accidentally leaving a blank, older versions you can roll back to, and checks that the finished dish matches the order before it leaves the kitchen.

When to reach for it

Reach for it when prompts are first-class assets: you need to substitute variables safely, lint a template before rendering, compose named sections (RTCE, critique, judge), enforce an output contract, or run an A/B experiment between prompt versions. If you just need a one-off string, plain template literals are fine — you don't need this. If you want the type definitions for prompt records rather than the engine, those live in @weaveintel/core.

How to use it

import { createSafeTemplate, extractVariables } from '@weaveintel/prompts';

const tmpl = createSafeTemplate('Summarize {{topic}} for a {{audience}} reader.');

console.log(extractVariables(tmpl)); // ['topic', 'audience']

const rendered = tmpl.render({ topic: 'tax law', audience: 'lay' });
console.log(rendered); // "Summarize tax law for a lay reader."

What's in the box

• Templates — createTemplate, createSafeTemplate, extractVariables, renderWithOptions, renderStructuredPromptMessages.
• Registry & resolution — InMemoryPromptRegistry, PromptResolver, resolvePromptRecordForExecution.
• Composition — frameworks (FRAMEWORK_RTCE, FRAMEWORK_JUDGE, renderFramework) and fragments (resolveFragments, {{>key}} includes).
• Quality gates — lintPromptTemplate, validateContract (JSON/markdown/code/length contracts).
• Experiments — InMemoryExperimentStore, weightedSelect.
• Provider adapters — openAIAdapter, anthropicAdapter, resolveAdapter.
• Evaluation & optimization — evaluatePromptDatasetForRecord, runPromptOptimization.
• Translation — buildTranslatePrompt, protectNonTranslatable, verifyTranslation.

Single entry point: @weaveintel/prompts.

License

MIT.