@spintax/authoring-prompt

The canonical prompt for writing spintax templates with a language model — one prompt, shared by the Telegram bot, the n8n node, the playground and any pipeline that drafts templates with an LLM, so that a template drafted anywhere is drafted the same way.

It is product content, deliberately not part of @spintax/core: the engine must not grow authoring opinions. The engine is a peer dependency for exactly one reason — the prompt asks it how many plural forms a locale has, instead of keeping a copy of the table that can drift.

Install

npm i @spintax/authoring-prompt @spintax/core

Use it

The loop every consumer runs: build the prompt, call a model, strip what the model wrapped its answer in, validate with the engine, and — when validation fails — ask the model to repair the exact spans the engine reported.

import { buildAuthoringPrompt, buildRepairPrompt, cleanModelTemplate } from '@spintax/authoring-prompt';
import { validate } from '@spintax/core';

const prompt = buildAuthoringPrompt({
  brief: 'A two-sentence welcome message for a new subscriber',
  locale: 'en',
  channel: 'email',
  allowedVariables: ['firstName', { name: 'product', note: 'brand name — do not inflect' }],
});

// Any model. `systemPrompt` is stable per locale and cacheable; `userPrompt` carries the brief
// and the per-request allow-list.
let template = cleanModelTemplate(await callModel(prompt.systemPrompt, prompt.userPrompt));

let diagnostics = validate(template, { locale: 'en' });
if (diagnostics.some((d) => d.severity === 'error')) {
  const repair = buildRepairPrompt(template, diagnostics, {
    locale: 'en',
    allowedVariables: prompt.allowedVariables,
  });
  template = cleanModelTemplate(await callModel(repair.systemPrompt, repair.userPrompt));
}

Pass the same locale to the prompt, to validate() and later to render(): it selects the grammar block and the plural arity the model is taught, and a mismatch is the bug the repair prompt exists to fix.

What it does not do

The prompt writes the prose of one block and never emits markdown or HTML; channel sets length and tone only (landing is a headline plus one sentence, not page markup), and no option relaxes that. If you need structure — an article of <h2> / <p> / <ul>, a document of sections — call the prompt once per block and compose the structure in your code: wrap each block in its tag, build the block-level permutations, write the separators yourself. That puts a whole class of template defects (a literal in a sep=, a dangling </p>, a triple newline from an empty branch) out of reach by construction. Want a longer block? Ask for it in brief — the channel is a default, the brief is the instruction.

API

• buildAuthoringPrompt(opts): BuiltPrompt — brief, optional locale, allowedVariables (bare names or { name, case?, note? }), channel, variationLevel.
• buildRepairPrompt(template, diagnostics, opts?): BuiltPrompt — takes validate() output verbatim; restate locale and allowedVariables so a repair cannot smuggle in a variable.
• cleanModelTemplate(raw): string — strips code fences, a template: label and wrapping quotes.
• promptExamples(locale?) — the worked examples the prompt teaches from, exported so a test can validate() the exact strings a model is shown.
• PROMPT_VERSION — bumps when the prompt text changes in a way that can change model output. Assert it in a pipeline that stores generated templates, and file conformance results under it.

BuiltPrompt is { systemPrompt, userPrompt, allowedVariables, promptVersion }.

Why it is shaped this way

The rationale is in docs/spec-llm-authoring-prompt.md. The reference loop — the one the committed conformance baseline was measured with — is conformance/run.mjs in this package's source.

License

MIT.