AI workflow · 15 posts
The model is 30% of the win.
I ship production features across a dozen repos with AI agents every day. The model matters far less than the harness around it — the verification gates, the deterministic hooks, the cheap always-on watchers, and the context architecture. This is the whole setup, warts and all, plus every way it has burned me.
The models and the money
Which model to point at the problem, what it costs, and why the answer keeps changing every quarter.
Model Review
Sol vs Fable 5: The Real-World Verdict After Two Weeks
GPT-5.6 Sol vs Claude Fable 5 after two weeks of production work: half the price, better in my hands, and why the benchmark split lied to me.
Read it →AI Economics
The Token Price Collapse: What Half-Price Frontier AI Actually Changes
Sol undercut Fable 5 by 50% and it rewires what you can afford to build. The batch maths, the cache maths, and the workloads that just became viable.
Read it →AI Workflow
Caveman Mode: Cutting AI Output Tokens 75% Without Losing Accuracy
Output tokens cost 6x input. Compressing how your AI talks back saves ~75% with zero accuracy loss — and the four cases where you must switch it off.
Read it →Opinion
10 Reasons Anthropic Might Be the Most Overvalued Company on the Planet
An opinion piece: Anthropic is priced like a monopoly-in-waiting while competing on commodity output, rented compute and a lead that resets quarterly.
Read it →How AI agents actually fail
Named failure modes from real sessions — unverified "done" claims, confident agreement, and building the wrong thing thoroughly.
AI Workflow
Verify Before Done: The Gate That Fixes 90% of AI Agent Failures
The single highest-leverage practice in AI-assisted development: never let the model's narration substitute for observing the real system. Here's the gate.
Read it →AI Workflow
10 Agentic Coding Failures I Hit With Claude That Sol Got Right
Ten specific agentic failures from one real session — assumed instead of asking, curl-200 as proof of done, over-building — and the fix for each.
Read it →AI Workflow
The Polite Liar: Why Your AI Agrees With You and Why That's Dangerous
Around 58% of AI responses show sycophancy, and agreeing with you increases hallucination. How to prompt for calibrated confidence instead of agreement.
Read it →AI Workflow
Scope Drift: Why AI Builds Four Things When You Asked for One
Over-building looks like effort and costs like a bug. Why agents route around the literal request, and the rules that keep them shipping exactly the ask.
Read it →AI Workflow
No Smoke Breaks: Why Your AI Agent Should Never Wait on a Timer
Fixed idle waits are pure latency — response time equals your polling interval. Event-driven agents beat polling 70-90%. When pacing is actually correct.
Read it →Building the harness
The model is maybe 30% of the win. This is the other 70% — context architecture, deterministic guardrails, cheap watchers, and multi-agent orchestration.
AI Workflow
How I Actually Use AI After Fable and Sol — My Optimised Setup
The harness, not the model: caveman token compression, cheap always-on log watchers, halls-and-rooms context files, deterministic hooks, and anti-footgun rules.
Read it →AI Workflow
Halls and Rooms: How to Structure AI Context Files That Don't Rot
A flat 4,000-line context file gets ignored. Split it into halls and rooms, load on demand, and treat every stale line as a bug. The full structure.
Read it →AI Workflow
Hooks Beat Prompts: Deterministic Guardrails for AI Agents
A prompt is a probability. A hook is a guarantee. Move anything you can't afford to get wrong out of your instructions and into the harness.
Read it →AI Workflow
Cheap Models as Always-On Watchers: The Pattern Nobody Uses
Stop paying flagship rates to stare at a log file. Put a small fast model on watch duty and build breaks become notifications instead of surprises.
Read it →AI Workflow
Running Multiple AI Agents Without Losing the Thread
Parallel sub-agents, worktree isolation, and agent-to-agent handover as precise tickets. What actually parallelises, and what a barrier really costs you.
Read it →AI Workflow
I Started Grading My AI Every Day. The Results Were Ugly
A one-row-per-day CSV with two counts and a letter grade turned vague AI frustration into fixable patterns. Six days of real grades: D, D, B-, E-, B, D+.
Read it →Let us make some quick suggestions?
Unique solutions for your brand.