Writing
What we have learned building AI agents, retrieval systems, automation pipelines and production software. Specifics over abstractions — including the parts that did not work.
If a note raises a question about your own stack, tell us what you are building.
17 notes shown.
An injection is not a jailbreak and it is not a prompt problem: a model enforces no boundary between the instructions it should obey and the data it should only read. The fix is an access-control one — break the capability combination, resolve the plan before exposure, carry provenance in code, and treat the renderer as an egress boundary.
A cap that raises an exception throws away everything the run learned; a cap reported as success hides the gap. Where the step counter belongs, how to see a loop before it reaches the ceiling, and what a truncated run owes the caller.
A test that fails one run in four and passes on re-run is usually a defect with a working alarm. Retries hide the evidence, quarantine files it — and the classification decides which fix you actually need.
A worker restarts mid-run and the invoice posts twice. Durability is not the model’s problem — it is the gap between doing an effect and recording it, and the three contracts that close it.
A non-deterministic system graded by a non-deterministic judge produces a check that either blocks every merge or is ignored. How we build eval gates that survive: deterministic checks as the gate, a measured noise floor, critical slices that block on a single case, and a judge kept advisory.
We design and build your site at no cost and host it free for three months. Here is the honest cost over three years, and where a free build is the wrong choice.
Most free CRMs cap seats or records, which is exactly what breaks a CRM. Here is why coverage matters, and what our free-forever tier does and does not include.
Review used to be a conversation between two people who had both thought about the problem. That contract is broken, and most teams have not replaced it with anything.
LLM cost is rarely one expensive thing. It is a small per-call cost multiplied by a call volume nobody modelled, with context that grows quadratically.
Nine agents on a diagram looked elegant and cost eleven times a single model call. Every agent boundary converts structured state to prose and back.
The blocker in legacy modernisation is never the typing. It is that nobody knows what the system does, so nobody can tell whether a change broke it.
Retrieve the chunks for a failing question and read them. Nine times out of ten the answer passage is simply not there — the generator was never the problem.
Route support tickets or approve supplier invoices. Same shape on a slide, completely different risk. Most automation programmes fail at selection, not build.
A model can take a module from 12% to 90% coverage before lunch, and the suite will catch nothing — because every assertion came from the implementation.
One client component in the root layout shipped 340KB the server could have rendered. Most App Router performance work is two or three structural decisions.
A missing await gets caught every time. A change that violates an invariant enforced three modules away gets approved. The difference is where the evidence lives.
The ticket still takes two hours. The time moved — into the review queue, and into the question the ticket asserted and never justified. Not acceleration. Redistribution.