Writing
Ideas worth
keeping.
Field notes on applied AI, organisational intelligence and the systems that make frontier models dependable. 54 essays, newest first.
Start here
Three ideas that frame the work.
-
Prompt, context, harness: three layers that accumulate.
A useful working model for AI practice: prompts define the job, context supplies the evidence and a harness makes both repeatable, even as models improve.
-
Central intelligence: governed memory a company can reason with.
Central intelligence connects organisational memory to current decisions with source evidence, accountable owners, permissions and review triggers.
-
A company brain should remember, cite and forget.
A company brain is governed operating memory: provenance, permissions, review triggers, deletion and a named human at the authority boundary.
The full catalogue
All writing, newest first.
- Software engineering, explained
Software Engineering
August 2026 - Machine learning, explained
Machine Learning
August 2026 - A working map
AI Engineering
August 2026 - AI security
An AI found a WordPress zero-day for $25. Read the fine print.
July 2026 - AI security
The closing cyber gap: open-weight AI is four to seven months behind the frontier
July 2026 - Governance
WAICO: China launches a global AI governance body - what it can and can't do
July 2026 - Release analysis
Claude Fable 5 and Mythos 5: what launched, what broke, and what it costs
July 2026 - Explainer
Mythos-class, explained: why Anthropic shipped the same model twice
July 2026 - Buyer's guide
Should you upgrade from Opus 4.8 to Fable 5?
July 2026 - Release analysis
Sonnet 5: Anthropic's cheaper way to run agents
July 2026 - Release analysis
GPT-5.6: three tiers, and a benchmark-gaming problem
July 2026 - Release analysis
Kimi K3: the largest open-weight model lands within reach of the frontier
July 2026 - Comparison
Which model should you code with in 2026: Fable 5, GPT-5.6 Sol, or Kimi K3?
July 2026 - Rumour watch
Gemini 3.5 Pro: what's confirmed and what's just rumour
July 2026 - Standards
MCP's biggest revision: what the 2026-07-28 spec changes
July 2026 - Regulation
The EU AI Act clock: what actually changes on 2 August 2026
July 2026 - Regulation
The Carrefour ruling: France sets the EAA enforcement precedent
July 2026 - Regulation
The EAA exemptions that don't save you: microenterprise and the 2030 runway
July 2026 - Operating model
When shipping gets cheaper, judgement and ownership matter more.
June 2026 - Infrastructure
What a data centre in 100 backyards must prove.
May 2026 - Release analysis
Opus 4.8: what the release benchmarks actually showed.
May 2026 - Markets
Anthropic raised $65B. What the primary record actually says.
May 2026 - /plan
Plan on paper, fail on paper.
May 2026 - Enterprise
The prototype is quick. Production is the work.
May 2026 - Architecture
The writer cannot be the only judge.
May 2026 - /execute
A plan you don't check is just a wish.
May 2026 - /brainstorming
Clarify consequential choices before you build.
May 2026 - /test
Write the test first, especially when the AI writes the code.
May 2026 - Agents
Agentic structure: how an answer becomes a controlled action.
May 2026 - /debug
Find the cause before choosing the fix.
May 2026 - Strategy
Rent the capability. Retain the learning.
May 2026 - /caveman
Caveman mode: when shorter output actually helps.
May 2026 - Context
Context engineering: what belongs in the window.
May 2026 - /verify
"Done" is a claim. Prove it.
May 2026 - Context
The context window is capacity, not a quality guarantee.
May 2026 - /finish
Finishing is a decision, not a Git push.
May 2026 - Harness
A better doer needs a better harness.
April 2026 - Craft
Why two people on the same model get different results.
April 2026 - /request-review
Request review while the decision is still reversible.
April 2026 - Prompts
Prompt engineering didn't die. It got absorbed.
April 2026 - /receive-review
Receive review: verify, then act.
April 2026 - Operating model
A context-window lens for organisational memory.
April 2026 - /compact
What Claude Code compaction keeps, loses and cannot prove.
April 2026 - Operating model
The company brain, honestly.
April 2026 - /subagent
Delegate bounded work, verify what returns.
April 2026 - Method
Find the pain a leader can't quite name.
April 2026 - /parallel
Parallel agents are fast only when the work can separate.
April 2026 - Field notes
What I'd tell a team starting with AI today.
April 2026 - /worktree
Git worktrees isolate local state, not design.
April 2026 - /write-skill
Turn repeated instructions into a skill you can test.
April 2026 - /use-skills
The first skill is choosing when a method applies.
April 2026