Writing Tools and AI: What Excites Me About the Intersection

Back in April, I wrote about what LLMs could mean for WriteDaily’s writing analytics. Since then, I’ve been running small experiments — not building product features, but understanding the technology.

What I’ve Been Testing

Local models. The privacy challenge with LLM-based analysis is the external API call. But local models are getting smaller and faster. I’ve tested several quantised models running on consumer hardware:

• Llama 2 7B (Q4) — Runs acceptably on an M1 Mac with 16GB RAM. Analysis of a 750-word entry takes ~3 seconds. The quality is surprisingly good for pattern recognition tasks.
• Mistral 7B — Faster than Llama 2, comparable analysis quality. The smaller footprint makes it more practical for a potential server deployment.
• Phi-2 — Much smaller (2.7B parameters). Analysis quality drops noticeably on subtle tonal shifts, but it’s fast enough for near-real-time feedback.

The bottom line: local LLM analysis is technically feasible. It wouldn’t require an external API, preserving WriteDaily’s privacy model. But it would require significantly more server resources than the current LIWC pipeline.

Prompt engineering for writing analysis. Getting an LLM to produce structured, consistent writing analysis is harder than I expected. Without careful prompting, the output drifts — one analysis is poetic, the next is clinical, the third invents patterns that aren’t in the text. My current best approach uses a two-pass system: first pass extracts structured data (word frequency, sentence variety, pronoun ratios), second pass generates narrative feedback constrained to the extracted data. The hallucination rate drops dramatically when the model isn’t asked to analyse and narrate in one step.

What I’m Not Doing

I’m not adding AI features to WriteDaily right now. Three reasons:

1. The current analytics are good enough. LIWC gives users meaningful insight into their writing patterns. Adding AI wouldn’t replace LIWC — it would augment it. The marginal value needs to justify the complexity.

2. The technology is moving fast. What’s the latest today is deprecated next quarter. Committing WriteDaily to a specific model or API would create maintenance burden I’m not ready to take on.

3. Privacy remains paramount. Until I can offer AI-powered analysis that runs entirely on-device or on WriteDaily’s own infrastructure, with no external data flow, it doesn’t ship.

Where This Is Heading

I’m continuing to experiment. The local model trajectory is promising — models are getting smaller, faster, and more capable. If the trend holds, running a capable writing analysis model on WriteDaily’s server (or even in the user’s browser via WebAssembly) will be practical within a year or two.

When that happens, I’ll be ready. But I’m not going to ship something that compromises the privacy and reliability WriteDaily’s users have trusted for a decade.

For now, writedaily.co continues to offer the same LIWC-powered analytics it always has. The AI experiments are happening in parallel, with no deadline and no promises.