A working journal from inside a technical writing team adopting AI faster than the org chart can keep up. Longform essays, tooling teardowns, and uncomfortable observations.
How LLM training actually works, shown in a working miniature transformer-style model with 44 knobs that you can train in your browser. Tokens, embeddings, Q K V attention, logits, loss, backpropagation and Adam, with every number traced...
Your agent ends every session with the same confident word. Here is the loop that checks that word from outside: a written definition of done, a boring checker, a heartbeat, and a grade. Five small files, one afternoon, real output included.
A full ITR-2 with capital gains, filed on incometax.gov.in by an AI agent for AY 2026-27, with no chartered accountant. The exact journey, the real gotchas, and how you repeat it.
You wrapped your documentation pipeline in Claude skills and wrote rules in markdown to keep them honest. Those rules are a sticky note, not a guardrail. Here is how a technical writer builds trust they can actually rely on into their own AI workflow: a check the agent can't override, provenance on every draft, and the seven things you do this week.
Thirty sources. Three hundred articles a week. Python fetches them and stays stupid on purpose. Claude reads everything and tells me the three things to do. A signal bank remembers the rest until it matters.
A frontier model got pulled three days after launch. The agent on your own machine has no off switch like that, so you cage it instead. A plain look at what NVIDIA's open-source OpenShell actually does.
How enterprise economics are compressing the craft.
Five research papers, a holiday week, and what I kept thinking about after I published the last one
Before someone else does it for you, and calls it progress
Your job title might say 'writer' but your skills better say 'architect, engineer, and occasional therapist.'
2025 taught us new tools. It also taught us what can't be automated away.
Google released Antigravity IDE this morning alongside Gemini 3. I spent my entire day testing it. The results surprised me.
The conversation with Raj that changed everything, the app I built for myself (not to sell), and an invitation to spend the next 365 days becoming the person AI can't replace
A hands-on weekend testing Google's new AI arsenal: Gemini Search, Jules, Stitch, and more—from a solo technical writer's perspective
A fine-tuned model trained on thousands of real documentation edits—rewriting with near-human precision