PromptCoach

A full-stack, AI-powered coaching tool that helps people write better prompts.

Most people writing prompts for AI tools never get feedback on why a prompt worked or didn't. They just see the output and guess. PromptCoach closes that loop. Paste a prompt in, and it's scored against a real rubric, explained in plain language, and rewritten to fix its specific weaknesses. Beyond one-off feedback, it includes a Learn section teaching the underlying principles and a Practice mode with a bank of realistic scenarios to apply them against, turning prompt engineering from guesswork into a skill you can deliberately practice and measure.

What it does

• Rate my prompt: paste any prompt and get an instant, structured evaluation: a weighted 100-point score, a plain-language verdict, and a rewritten version that addresses the specific gaps found. If the prompt is too vague to evaluate honestly, it asks one clarifying question instead of guessing, and never asks twice.
• Learn: short, skimmable lessons, one per rubric concept, each with a real bad-prompt/good-prompt pair and a two-to-three sentence explanation of why the difference matters.
• Practice: a bank of ten realistic scenarios (data extraction, code generation, summarization, classification, and more) graded against that specific task's requirements, culminating in a deliberately underspecified "trap" scenario that tests whether you've internalized the habit of adding your own clarity, not just following instructions.

Technology stack

Layer	Choice	Why
Frontend	React (Vite)	Fast local dev, small production bundle, no framework overhead for a focused single-purpose app
Hosting	Azure Static Web Apps	Serves the frontend and links directly to a managed Functions backend: one deployment pipeline, near-zero idle cost
Backend	Azure Functions (Node.js)	Serverless HTTP endpoints; scales to zero when idle, scales out automatically under load
AI	Anthropic Claude API	Structured, tool-forced output (not fragile text parsing) drives every rating, so the frontend always gets predictable, typed JSON back
Auth	Azure Static Web Apps built-in auth (GitHub)	Zero custom auth code or session management: the platform handles the OAuth flow and identity headers natively
CI/CD	GitHub Actions (auto-provisioned by Azure)	Every push to main builds and deploys automatically

Engineering highlights

A scoring rubric refined through empirical testing, not guesswork. The rating system went through several full iterations, each driven by deliberately adversarial test prompts designed to expose specific failure modes: a prompt with genuinely conflicting requirements, a prompt requiring clarification, a wall-of-text prompt with no visual structure. Each round surfaced a real, specific miscalibration (scores that didn't match their own written justification, a model second-guessing plainly-computed facts), which was then fixed and re-verified against the exact prompt that exposed it, not just patched and assumed fixed.

Deterministic math where it matters. The final score is a weighted combination of five dimensions, but the weighting arithmetic happens in application code, never inside the AI's own response, removing an entire category of run-to-run inconsistency that a purely model-computed score would carry. The same principle applies to structural analysis (line breaks, word count): those are computed directly from the text in code and handed to the model as established facts, rather than trusting the model to notice formatting on its own.

One scoring engine, reused everywhere. Practice mode's task-specific grading reuses the exact same rubric and scoring pipeline as the general-purpose rater, rather than a second parallel system. A task's specific requirements are folded into the same evaluation as additional context, so scores stay directly comparable across every part of the app.

Built with an AI-assisted engineering workflow. PromptCoach itself was built in close collaboration with Claude Code, with Dylan directing architecture decisions, designing and iterating the scoring rubric, and verifying every change against a live, running instance before considering it done, a fitting way to build a tool about writing better AI prompts.

Status

Actively developed. Core features (rating, Learn, Practice mode, GitHub login) are live; API rate limiting and cost-control infrastructure are in progress ahead of public launch.

---

See also this platform and Plot Twist.