Thinkings
Thinkings.
Notes from the questions behind our work: context, retrieval, efficiency, agents and the systems around them. Some are measurements, some are working hypotheses, and some document what did not work.
We publish these while we are still working on the questions, which means a few of them will age badly. That seems better than only publishing conclusions — a note that turned out to be wrong is usually the more useful one to have written down.
None of it comes from client work. What we learn with a company stays with that company; what is published here comes from our own products, our own experiments and our own reading.
The archive
Archive.
Newest first. Efficiency notes tend to carry numbers and the assumptions behind them; retrieval and systems notes are more often arguments. Both are written to be checked rather than agreed with.
Every answer carries three prices: francs, seconds and watts. All three move with the same number, and the third appears in no dashboard.
Tokens behave like money. Unbudgeted, they get spent without anyone noticing. Why a longer window is not free, and how more context can make answers worse.
Search finds what is similar. Retrieval has to decide what a task needs. On embeddings, how documents get cut up, and the metric that misleads.
Why we write the question down before we write code. With an example that did not work, and the reason it still sits on the shelf.
An answer with no trace to a source is a claim. Why footnotes added afterwards make it worse, and what traceability does not give you.
Reuse is not an optimisation but a stance. How a prompt cache actually works, why a timestamp destroys it, and what it costs when you place it wrong.
A tool that does not know its own context hands the thinking back to the human. Three design rules from the first year of LeanCTX, including the exception to the second.