PROMPTSMITH: the prompt forge

Raw layman input in. Master-grade prompt out.

A free, open-source prompt compiler. Describe what you want in plain English and PROMPTSMITH compiles it into an expert-grade, structured prompt that an AI (Claude, v0, Cursor, ChatGPT, etc.) executes far better than a casually-typed request. It also bakes in an anti-AI-slop ruleset (or, for code, a production-WordPress ruleset) so the output looks human-made, not machine-stamped. It runs 100% free with no API key, and it learns from your feedback so every prompt is sharper than the last.

Built for an agency, open-sourced for everyone. MIT licensed: use it, fork it, deploy it.

Why it exists

"Vibe coding" gives garbage-in → garbage-out. A loose prompt produces the telltale AI look: Inter/Space Grotesk fonts, purple-on-white gradients, the centered-hero + 3-card-grid template, lorem ipsum. PROMPTSMITH fixes the input so the output is sharp, specific, and on-brand every time, regardless of who's prompting.

Domains (v2)

PROMPTSMITH compiles prompts for four domains, across two pipelines:

Domain	Pipeline	What it produces
Frontend / Design	design	Landing pages, sections, wireframes, redesigns (HTML/React/Next/Tailwind/Vue)
Elementor	design	Page templates, sections, headers/footers, popups (a build recipe + Custom CSS)
Custom Widget	engineering	Elementor widgets, Gutenberg blocks, shortcodes (production WP code)
WP / Woo Plugin	engineering	WordPress & WooCommerce plugins (secure, standards-compliant code)

• The design pipeline runs the aesthetic-direction + anti-AI-slop + human-feel engine.
• The engineering pipeline swaps those for WordPress coding standards, security (sanitize / escape / nonces / capabilities / $wpdb->prepare), i18n, and an anti-boilerplate code-quality ruleset, and auto-injects a WooCommerce block when the brief touches Woo.

Adding a domain pack is mostly data: extend lib/knowledge/domains.ts (kinds + targets); the compiler routes by pipeline automatically.

How it works (the hybrid engine)

1. Deterministic compiler (lib/compiler.ts): 100% free, no API key, instant. Takes the structured intake and assembles a prompt using a CO-STAR-derived structure (Role · Context · Objective · Aesthetic · Constraints · Output) fused with the frontend-design discipline. This alone produces a ~1,000-word expert prompt.

2. AI Boost (optional, app/api/boost/route.ts): a "Boost" button hands the compiled prompt to an AI that sharpens it without weakening any constraint. It runs on a multi-provider failover chain (Groq → OpenRouter → Claude) so it stays up on free tiers: when one provider rate-limits or errors, the next takes over automatically. Works fully without any key; Boost just adds polish.

3. The learning loop. The system gets better the more it's used:

   • Every forge auto-saves to a shared server-side store (data/promptsmith.json).
   • After using a prompt, devs submit feedback: outcome (nailed / partly / missed), structured issue tags (e.g. "used a generic font", "fell back to the template layout"), and plain-language notes.
   • Failures distill into reinforced "lessons" (lib/learning.ts), scoped by project-type + framework, with a weight that grows as the same issue recurs.
   • Future prompts in that category auto-inject the relevant lessons as a high-priority block, and since AI Boost reads the whole prompt, Claude inherits them too. No model training; just feedback-driven rule reinforcement that compounds.
   • The Memory tab shows accumulated lessons, the success rate, and forge history.

The real IP is in lib/knowledge/. Edit these to tune your agency's house style:

File	What it controls
knowledge/antiSlop.ts	The bans (fonts, gradients, template layouts) + positive forcing functions
knowledge/humanFeel.ts	Rules that make output read as human-crafted
knowledge/aesthetics.ts	7 bold aesthetic directions + font pairings, auto-selected from the "vibe"
knowledge/frameworks.ts	Per-target tech constraints + output format (HTML, React, Next, Tailwind, Vue, Elementor)
knowledge/projectTypes.ts	How each deliverable (landing / section / wireframe / redesign) is framed
knowledge/domains.ts	The domain registry: 4 packs, each with kinds + targets + pipeline
knowledge/wpEngineering.ts	WP standards, security, code-quality, WooCommerce rulesets (engineering pipeline)
learning.ts	Maps feedback issue-tags → reinforcement rules; distills + selects lessons (domain-scoped)
store.ts	Shared JSON store (sessions / feedback / lessons), atomic writes, swappable for SQLite/Postgres

Features

Forge view

• Plain-English brief → expert prompt, with deliverable + build-target + density controls
• Direction override: auto-selects an aesthetic direction from your vibe, or force one of 7
• Avoid field: brief-specific negative constraints (hard bans)
• Live readiness signals + a count of learned lessons that will be injected
• ⌘↵ / Ctrl+↵ to forge; Copy, download .md, AI Boost, ↗ Claude / ↗ ChatGPT (copies + opens), Breakdown, Recompile
• Draft auto-saves to localStorage; honest header chip shows whether AI Boost is configured

Memory view (the growing brain)

• Stats: prompts forged · feedback logged · lessons learned · success rate
• Lessons: auto-distilled from feedback, plus add your own house rules, mute, or delete any lesson
• ✦ Distill notes: (with a key) Claude reads freeform feedback and proposes new lessons
• Recent forges: click any to reload its full brief back into the editor

Run it

npm install
npm run dev          # http://localhost:3000

Optional AI Boost:

cp .env.example .env
# add ANTHROPIC_API_KEY=sk-ant-...

Deploy your own (free)

One-click on Vercel, or:

npm i -g vercel
vercel            # preview
vercel --prod     # production

It deploys with zero config and no env vars: the free deterministic engine works out of the box. For a public instance, mind these:

Env var	Why
GROQ_API_KEY / OPENROUTER_API_KEY / ANTHROPIC_API_KEY	Enable AI Boost + AI-distill. Set any/all; they form a failover chain (tried in order; when one rate-limits or errors, the next takes over). Groq + OpenRouter both have free tiers → free, resilient AI for your devs.
GROQ_MODEL / OPENROUTER_MODEL / PROMPTSMITH_BOOST_MODEL	Optional model overrides (good free-tier defaults built in).
PROMPTSMITH_AI_ORDER	Failover order, e.g. groq,openrouter,anthropic (default).
PROMPTSMITH_ADMIN_TOKEN	Set on any public deploy. Locks store-mutating endpoints (add/mute/delete lessons, AI-distill) behind an x-promptsmith-admin header. AI Boost stays open for users.
NEXT_PUBLIC_SITE_URL	Your deploy URL, for OG/social share cards.

Persistence note: on serverless (Vercel), the JSON store falls back to the temp dir, so the learned "brain" is per-instance and ephemeral (the app detects this and shows a banner). The prompt compiler is fully stateless and works perfectly. For a durable, shared brain, self-host on a persistent server, or swap lib/store.ts for KV/Postgres (the module's API is the only surface callers depend on).

License

MIT © Systical. Free for any use, including commercial. No warranty.

Tech stack

• Next.js 15 (App Router, TypeScript): standalone web app + the Boost API route
• Bespoke CSS (app/globals.css): no Tailwind; full control over the obsidian-and-ember precision-instrument aesthetic (deliberately not AI-slop)
• @anthropic-ai/sdk: optional, only used by the Boost route (system prompt is cached)

Roadmap

• v1: frontend / design prompts (landing, section, wireframe, redesign). ✓
• v1.1: direction override, avoid field, lesson management, AI-distill, history reload, QoL. ✓
• v2 (this build): domain packs (Elementor, Custom Widget, WP/Woo Plugin), with a full engineering pipeline (WP standards / security / i18n / code-quality / WooCommerce). ✓
• Next: per-dev house-style profiles, prompt presets, a Gutenberg full-site-editing pack, and optional SQLite/Postgres for the store.