America/Bogota
blog // post
session
All posts
2026-09-21

Checklist before you build the bot

Most bots fail because the process was never ready for a bot. A short checklist, before agents, queues, or clever prompts, saves you from automating chaos.
2026-09-21
Before I touch an agent, a queue, or a clever prompt, I want a boring checklist. Not because bots are bad. Because most “AI projects” I see are a process problem wearing a model costume. This sits next to human-in-the-loop is the product: HITL is what you keep. The checklist is what you do before you deserve a bot at all.
  1. Name the job in one sentence. Not “improve ops.” Something like “get a candidate from uploaded docs to ready-for-affiliation decision.”
  2. Write the kill condition. Falsifiable automation: I would turn this off if _____. If you cannot write it, you are not ready to turn it on.
  3. Map the human judgment points. Where must a person decide — and what evidence do they need on screen? Those points become product surfaces, not afterthoughts.
  4. Count context hops today. Portals, chats, spreadsheets, email. If the bot does not remove hops, it only moves typing.
  5. Define the receipt. What proof will operators trust after a run (counts, IDs, timestamps)? Without a receipt you buy trust latency on day one.
  6. Cap blast radius. One bad dependency should not freeze the whole hiring week. Partial progress beats all-or-nothing magic.
  7. Ship the smallest loop that can die honestly. One queue, one document type, one happy path with a loud fail — then expand.
  • A flow nobody can narrate end to end.
  • A “bot” whose only success metric is “messages sent.”
  • Anything whose failure mode is silent and looks like success in the UI.
  • Work where the expensive part is judgment and the cheap part is typing — unless the bot clearly feeds the judgment, not replaces it with confidence theater.
On tools.danieldiaz.dev the blast radius is usually me. On ECORPA it is someone else’s hiring timeline. The checklist does not change shape. The cost of skipping a line does. If the list feels heavy, that is the point. It is still cheaper than a bot that operators learn to walk around within a week — which is just another drawer entry, except this one burns tokens and trust.