At Nurture Boss, reading real conversations showed the assistant botched relative dates like "two weeks from now" 66% of the time. Targeted tests for that one class took the success rate from 33% to 95%.
The bottom-up method: write open-ended notes on bad traces, have an LLMA large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.Full definition → cluster the notes into a failure taxonomy, then count frequencies. Three failure modes accounted for over 60% of all problems.
Generic dashboards do damage twice: teams feel data-driven while tracking numbers no user cares about, and optimizing many dimensions at once means none of them get attentionThe mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.Full definition →.
Prompts are just English, so routing a learning designer's expertise through slides and into an engineer's prompt adds a translation step for nothing. Give domain experts an admin UI that edits the live prompt.
Build the annotation tool for your domain: all context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → on one screen, one-click correct/incorrect, hotkeys to move between examples. Every extra click is why systematic review quietly stops happening.
Terms in this piece · Glossary
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
Why it matters
If your AI team obsesses over frameworks and vector DBs but can't tell whether changes actually help, this lays out a measurement-first workflow — error analysis, lightweight data viewers, synthetic data, and roadmaps that count experiments not features — drawn from 30+ real deployments.
Key quotes
“Teams with thoughtfully designed data viewers iterate 10x faster than those without them.”
Hamel
“LLMs are surprisingly good at generating excellent - and diverse - examples of user prompts. This can be relevant for powering application features, and sneakily, for building Evals. If this sounds a bit like the Large Language Snake is eating its tail, I was just as surprised as you! All I can say is: it works, ship it.”
Bryan Bischof
“Seeing how the LLM breaks down its reasoning made me realize I wasn’t being consistent about how I judged certain edge cases.”
Phillip Carter
“The key metric for AI roadmaps isn’t features shipped – it’s experiments run.”
