Salsa for the bodega: how to make your estimates actually useful

Salvador has been over-estimating for a while now. The pattern is simple: LLMs default to padding, and every small "round up just in case" decision compounds until the numbers stop meaning anything. He's hit it enough times that he now reads "months" as a signal to discount — it's a lower-bound, not a ceiling.

His fix is to write estimates like a person who actually knows what they're doing. When Claude is the engineer and Salvador is steering, things are faster than the default. Days are for things that genuinely take days — real ML training runs, multi-protocol design, sandbox infrastructure — not for anything more than thirty lines of code. If something is multi-week, say so. If unsure, give the low number first and only offer a higher bound when there's a real reason.

The piece is worth a read because it's about the craft of explaining things clearly, not about any specific tool. The rules apply whether you're writing prompts, scoping features, or telling someone how long a project will take. The lesson is the same: stop hedging. Give a concrete number.

Why this matters for us: we're writing for the same people — busy, sharp, tired of the waffle. Brevity with conviction is the voice we're building.