We live in an age where AI can rewrite anything. Hit a button, and your clunky draft becomes polished prose. But here’s the thing - that doesn’t make you a better writer. It just makes you dependent on AI.

What if instead of rewriting your work, AI could coach you to improve it yourself?

That’s the idea behind The Writing Gym, a prototype I built in a focused evening session. Rather than offering quick fixes, it creates a structured reflection loop that helps writers identify their blind spots and build genuine skills.

The Core Insight: Coach, Don’t Rewrite

The philosophy is simple but powerful. Instead of letting AI do the work for you, the tool guides you through a process:

1. Context Setting - Answer pre-draft questions about your audience, intent, and concerns
2. Criteria Extraction - AI analyzes your context to identify what makes good writing for your specific situation
3. Draft & Self-Assess - Write your piece, then rate yourself on each criterion
4. AI Assessment - Get an independent AI evaluation with detailed reasoning
5. Comparison - See where you over/underestimated yourself and learn from the gaps

The magic happens in that comparison step. When you consistently overestimate your clarity but nail your tone, you start to recognise patterns in your writing blind spots.

Building the Flow: From Idea to Working Prototype

The Technical Foundation

I went with a clean, familiar stack: React + Vite + Tailwind for the frontend, Express for the backend, with Anthropic’s Claude handling the AI heavy lifting. Sometimes boring tech choices are the right tech choices - I wanted to focus on the core experience, not wrestle with new frameworks.

The app architecture is essentially a state machine moving through five steps, with API calls to extract criteria and assess drafts. Simple, predictable, debuggable.

The Pre-Draft Questions Component

This felt like the most crucial piece to get right. The questions needed to be coaching-focused, not just data collection:

• Who is your audience?
• What do you want them to think, feel, or do?
• What are you most concerned about with this piece?
• What type of writing is this?

These aren’t arbitrary - they mirror what experienced writers naturally consider before drafting. The AI uses these answers to extract personalised criteria, turning generic writing advice into specific, contextual guidance.

The Assessment Loop

Getting the AI assessment right took some prompt engineering. I needed Claude to:

• Score each criterion consistently on a 1-5 scale
• Provide specific reasoning tied to the draft content
• Stay focused on coaching rather than rewriting

The breakthrough was including the full context (audience, intent, concerns) in the assessment prompt. This gave Claude the same information the writer had, making its evaluation more relevant and actionable.

The Split-View Revelation

Initially, I built this as a linear wizard - complete one step, move to the next. But during testing, I realised writers need to iterate. You don’t just assess once and move on; you revise and reassess multiple times.

The solution was a split-view workspace: draft editor on the left, scrollable iteration history on the right. Suddenly it felt like a real writing tool, not just a one-time assessment. This architectural change transformed the prototype from “interesting demo” to “something I’d actually use.”

The Wins and Surprises

The comparison view is addictive. Seeing “You rated this a 4, AI rated it a 2” with specific reasoning creates genuine learning moments. It’s like having a writing coach look over your shoulder.

Context matters more than I expected. Generic writing criteria are useless. “Be clear” means nothing. “Be clear enough that a busy executive can scan this and understand the key decision points” is actionable.

The tool stays in its lane. By focusing on assessment and reflection rather than generation, it avoids the “AI doing your homework” problem. You still have to write; you just get better at evaluating and improving your own work.

What I Learned

Coaching tools have different design constraints than productivity tools. The goal isn’t efficiency - it’s skill development. That means embracing some friction, like making users reflect before getting AI feedback.

AI, like a good Scrum Master, works better as a mirror than a magic wand. When Claude assesses your draft against your stated intentions, it creates accountability. You can’t blame the AI for missing your point if you never articulated it clearly.

State machines are underrated for complex user flows. Having explicit steps made the app predictable and easy to reason about, both for users and developers.

What’s Next

This was a focused prototype session—about 2.5 hours from idea to working demo. The core loop is solid, but there’s room to grow:

• Pattern recognition: Track blind spots over time and surface trends
• Criteria library: Build up common criteria patterns for different writing contexts
• Collaborative features: Let writing groups share and discuss assessments
• Integration: Connect with existing writing tools and workflows
• Mobile Decisions: How does this work on a small screen? Is extracting the success criteria enough?

The code is open source under AGPL-3.0 if you want to experiment with it or build something similar.

The Bigger Picture

We’re at a crossroads with AI writing tools. We can use them as crutches that make us dependent, or as coaches that make us stronger. The Writing Gym is my attempt to explore that second path.

Because at the end of the day, being a good writer isn’t about having access to the best AI rewriter. It’s about understanding your audience, recognising your blind spots, and improving your craft through deliberate practice.

Sometimes the best way to get better at something is to get better at evaluating it first.