AI
Practical guides for developers working with AI tools. Prompt engineering, custom instructions, and integration patterns for ChatGPT, Claude, and other LLMs.
The Agent Skills That Pay Off Are Scripts With a Manual Attached
Curated skills add 16.6pp, agent-written ones add nothing, and half never trigger. What makes a SKILL.md pay off, and how to vet third-party skills.
Your CLAUDE.md Describes Your Folders. The Agent Already Knows Them.
ETH Zurich found repo overviews in context files don't help agents find code. What belongs in CLAUDE.md, memory, skills and hooks instead.
jevmem Writes Your Project Memory for You. Read Where the Turns Go First.
jevmem auto-saves Claude Code decisions to JEVMEM.md. How it works, the install steps, and what it sends to TypeSafe AI - for solo and enterprise use.
I Shipped Five Prefilled AI Links, Then Read the Reprompt Write-Up
Reprompt turned Copilot's ?q= parameter into an exfiltration channel. What that means if you ship share-to-AI buttons, and the five URLs I actually use.
Prompt, Context, Loop, Graph: Every AI Discipline Is a Patch With an Expiry Date
Each discipline patched one weakness and expired when the weights caught up. The half-life table, the eight-week naming cadence, and where graphs actually fit.
Your Agent Isn't Slow, It's Queued
Half the wall clock in a typical agent pipeline is false sequencing. Same work: 130s as a chain, 65s as a fan. Here is the arithmetic, plus the code.
Claude Now Watermarks Everything It Writes, and There Is No Flag to Turn It Off
Anthropic swapped the sampler's randomness for a keyed choice. The mechanism, the length threshold, what survives editing, and why there is no opt-out.
Where AI Actually Went in 2026: Agents, Context, and the Quiet Wins
Past the hype: agents that ship, context engineering, MCP as a standard and multi-agent orchestration - what's real in 2026 and what's still a demo.
The AI Releases That Actually Mattered This June
An opinionated roundup of the June 2026 model releases - GPT-5.5, Gemini 3.5, Claude Sonnet 4.8, Mellum2 and Nemotron 3.5 - and which ones matter.
Six Months with Claude Code: The Workflow Habits That Cut My Token Spend
Six months with AI coding agents: when to delegate, how to manage context, and the habits that separate productive agent use from expensive frustration.
Model Context Protocol, Honestly: What MCP Solves and Where It Still Hurts
What MCP is, how its client-server architecture works, how it compares with function calling, and an honest assessment of where it still hurts.
Eight Prompt Engineering Patterns I Use Daily — and Four That Did Nothing
Structured output, chain-of-thought, few-shot examples, role framing and constraints - the prompt patterns that hold up, and four that did nothing.
AGENTS.md Makes Your AI Coding Agent Worse - and Now There's Research to Prove It
ETH Zurich tested AGENTS.md on four coding agents: no reliable gain in success rate and 20%+ higher cost. What the paper found, and what works.
The 4.5× Token Tax of Bloated AI Custom Instructions (and How to Cut It)
How custom instructions for ChatGPT, Claude and Cursor inflate every prompt, what that costs in tokens, and how a lean instruction set compares.
A Local RAG Chatbot for Your Internal Wiki: Ollama, ChromaDB, Docker, Zero API Keys
A local RAG assistant that indexes your docs with vector embeddings and answers from a local LLM. FastAPI backend, browser UI, full Docker Compose setup.
After a £47 OpenAI Bill on a Side Project, I Moved to Local LLMs. Here's the Stack.
Run Ollama in Docker Compose and build a Python CLI that clusters customer reviews by theme and writes a structured report. No API keys, no monthly bill.
Less Prompt Craft, More Context Engineering
The pattern across everything written here is that the wins came from managing context, not from clever wording. Eight prompt patterns is the practical starting point, and it is deliberately honest about the four that made no measurable difference. Six months with Claude Code covers the working habits - when to delegate a whole task and when to drive it turn by turn.
Custom instructions are where most people lose money without noticing. Every instruction is re-sent on every request, so a bloated set is a permanent tax rather than a one-off cost; the token tax breakdown puts real numbers on it. There is now research pointing the same way, covered in the AGENTS.md study, which found that bigger context files made agents worse rather than better. If you would rather start from something tested than a blank file, the custom instructions templates cover ChatGPT, Claude and Cursor.
Two builds go end to end rather than staying theoretical. The local feedback analyser replaces a per-token API bill with Ollama in Docker Compose and a Python CLI that clusters reviews by theme. The local RAG chatbot indexes your own documents with vector embeddings and answers from them, with a FastAPI backend and no API keys anywhere. Both run entirely on your machine.
For the wider picture, MCP explained covers the tool-integration protocol and where it still hurts, while the 2026 trends piece and the June release roundup sort what shipped from what demoed.