WorkAI Prompt Intelligence Platform
SaaS for the craft of instructing models
AI Prompt Intelligence Platform
A workspace that turns rough prompts into engineered ones — improvement, quality scoring, reusable templates and agent-loop design in one place.
- Status
- Live
- Category
- AI product
- Year
- 2026
- Role
- Solo build: product design, architecture, engineering, deployment.
Live Beta — deployed and evolving. No user or revenue claims, deliberately.
The problem
Prompt quality is the highest-leverage, least-tooled part of working with language models. Teams pass folk knowledge around in docs, lose their best prompts in chat history, and have no way to measure whether an 'improved' prompt actually improved.
Constraints
- A real SaaS surface from day one — auth, admin, audit trail — no demo shortcuts
- Provider-agnostic by design: the product must survive any single model vendor changing
- Honest beta posture: no invented usage numbers anywhere on the surface
The approach
Build the workspace I wanted: paste a rough prompt, get an engineered one back with a quality score and the reasoning; keep everything in a versioned library; and treat advanced patterns — especially agent loops — as first-class templates rather than tribal knowledge. Ship it as a real SaaS from day one: authentication, admin, audit trail, migrations — the unglamorous 80%.
What was built
- A prompt-improvement engine with structured quality scoring and category-specific rubrics
- Task-specific template system — including an agent-loop category covering per-iteration state, evaluation rubrics and exit conditions
- A reusable prompt library with version history
- Three AI-provider adapters behind a routing layer, so tasks run on the model that suits them
- Admin panel with an audit trail, on hardened Postgres row-level security
- Forty-seven automated tests across the engine and platform surface
Contribution
Built with Next.js · TypeScript · Supabase · Postgres RLS · Multi-provider AI adapters · Vercel
Technical decisions
Audit before growth
Admin tooling and an audit trail went in before any growth features. A platform that edits people's working prompts needs accountability from the first user, not the thousandth.
Agent loops as a first-class category
Most prompt tools stop at single-shot prompts. The loop template — state, rubric, exit conditions, failure handling — encodes how agentic work actually gets engineered.
Route by task, not loyalty
Different models earn different jobs. The adapter layer keeps the product honest about that instead of hard-wiring one vendor.
Evidence
Outcome
- Deployed and versioned as a working public beta (v0.1.x)
- Category set actively expanding — agent-loop engineering shipped as its own discipline
- A working demonstration that I can take a SaaS from zero to deployed: auth, billing-ready schema, admin, audit
Current status
LiveLive Beta — deployed, versioned (v0.1.x), forty-seven tests green; the category set is mid-expansion.
Lessons
The AI is the headline, but shipping a real SaaS surface — auth, admin, audit, migrations, tests — is most of the product. That's the part this build proved.
Next stage
A staged loop-engineering release ships next; billing and team workspaces follow validation rather than precede it.