10 issues so far
Writing
Longer pieces on putting AI to work on ordinary tasks. Every article is an issue, and every issue gets its own cover.
- A link that pings my phone when the work is doneA small relay that turns one link into a Telegram message. Why it answers before it delivers, why the template is mine and the text isn't, and how my AI code review uses it to tell me it's finished.
- When not to automate yetMost reasons not to automate are really reasons to wait: the rules are not agreed, the process is still changing, judgement is the job, or nobody will notice when it breaks.
- The manual-work audit: which tasks are worth automatingMost automation projects fail at the choosing, not in the code. A short audit that scores tasks by frequency, variance and the cost of a mistake, counts what manual work really costs, and turns the few that pass into a business case.
- An AI agent without the hype: a loop with limits"Agent" has come to mean almost anything. What an agent actually is, when a plain workflow is the better choice, and what makes one safe to leave running: limits in code, narrow tools, a log of every step, honest gaps and a person where mistakes cost money.
- From a one-off script to a tool people can rely onA script that saves an hour becomes a liability when it breaks and only its author understands it. Safe reruns, clear errors, config, logs and alerts, and a named owner: the plain engineering that turns it into a tool.
- Invoice extraction you can trust: check the numbersReading the fields off an invoice is the easy part. How to make the extracted numbers trustworthy: arithmetic checks with exact decimals, checks against your own records, routing by what failed, and counting exceptions instead of accuracy.
- A knowledge assistant that can say "I don't know"A chatbot over your documents is easy to build and easy to trust too much. How to make it cite its sources, refuse when it has none, respect who can see what and stay up to date, with running costs as of 25 September 2026.
- Text-to-SQL, safely: roles, row security and query checksA model that writes SQL is easy to demo and easy to get wrong. How to make it safe with column grants, row-level security and a query checker, so that even a manipulated model can only read what the user is allowed to see.
- A security checklist for AI featuresPrompt injection, tool permissions, personal data, output checks and spending limits: the controls that keep an AI feature safe without relying on the model to behave, mapped to the OWASP Top 10 for LLM applications.
- Choosing a model: price, speed and qualityWhy the most expensive model is rarely the right default. Claude API prices as of 25 September 2026, a worked example of a cheap-model-first setup, and how to test models on your own examples.