meta prompt architect lite

Turn a plain-language task description into a minimal prompt template — extracts {{variable}} placeholders, detects output format, assembles a ROLE/TASK/INPUT/CONSTRAINTS/OUTPUT scaffold. Use when the user asks to write or templatize a prompt for an LLM.

scripts/generate.mjs deterministically extracts structure from a task description; you refine wording and variable names. The script never invents requirements — it surfaces what's missing.

Rules

• The template is a scaffold, not the final prompt. Rewrite vague sections with the user's specifics.
• Keep {{variable}} names stable between INPUT and references.
• The CONSTRAINTS injection rule is load-bearing — do not remove it.
• Resolve each warnings[] item by asking the user or stating an explicit assumption.

1. Run

Requires Node 18+, zero dependencies. Fully offline.

bash

node <skill-dir>/scripts/generate.mjs task.txt
echo "..." | node <skill-dir>/scripts/generate.mjs -

JSON on stdout; exit 1 when warnings exist, 2 on read/usage error.

2. Read the report

• variables[]: {name, source} candidate {{placeholders}}.
• format: detected output format or null.
• warnings[]: gaps to close.
• template: the assembled scaffold.

3. Refine before delivering

1. Rename variables to the user's domain vocabulary.
2. Fill # OUTPUT with a concrete shape (JSON schema, section list, length cap).
3. Add domain constraints the description implies.

Lite vs full

The full meta-prompt-architect additionally detects tone and audience, adds assumption/ambiguity edge-case rules, and emits happy-path / edge / adversarial test-case stubs.