Generate unique product description pools with OpenAI and a Spintax node

Quick overview

Self-hosted n8n only — this template uses a community node. OpenAI writes one spintax template from your brief; the Spintax node turns it into a pool of unique, linted product description variants, validated and repaired once before rendering.

How it works Starts when you manually execute the workflow. Defines the writing brief and fixed product variables (product, brand, audience, feature). Builds an authoring prompt and sends it to OpenAI to generate a spintax template. Validates the returned spintax and, if it is invalid, generates a repair prompt, asks OpenAI for a corrected version once, and re-validates. Renders 12 description variants from the cleaned spintax template using a fixed seed and the provided variables. Lints each rendered variant to catch issues that only appear after rendering (for example repeated words or punctuation/spacing defects). Drops near-duplicate variants by similarity scoring and outputs the remaining pool as ready to publish.

Setup Add OpenAI credentials on "Chat model for authoring and repair" (or replace it with another compatible LLM) — that single node feeds both LLM chain steps. Edit the brief and the product variables in the “Your brief and product” step to match your item and style constraints. Adjust rendering and quality thresholds as needed (variant count, seed, lint window/ignore list, and the Jaccard dedup threshold) before activating the workflow.

Requirements Self-hosted n8n with community nodes enabled (instance admin setting) The n8n-nodes-spintax community node, version 0.2.1 or newer — install via Settings → Community Nodes An OpenAI credential, or any other LLM node you already use — the Spintax node itself needs none: its engine is bundled and renders locally, with no network calls

Customization Change the Base Seed on "Render 12 variants" to re-roll the whole pool — the same seed always returns the same twelve documents Put the strings every document repeats — product name, brand, a merge tag — into Lint's Ignored Strings and Uniqueness's Shared Strings, so both checks judge your writing and not your data Lower Footprint Limit on "Check pool uniqueness" from 1 to 0.15 to turn the footprint from a report into a gate that stops a pool sharing one skeleton

Additional info Why two checks and not one. Validate judges the template; it cannot judge the render. A flawless template still turns out the odd broken line, because that defect lives in the combination of choices rather than in the source — two neighbouring slots picking the same word, an unlucky join leaving a space before a comma. Lint reads each rendered document. Uniqueness reads the whole surviving pool and reports the share of five-word windows that repeat across it: measured on real pools of equal size, one template scores about 0.96 and six templates about 0.02, and asking for more variants of the same template cannot move that number — only new templates, or denser variation inside the one you have.