Make your Robots Argue

The decision is the easy part. You already knew you were going with Postgres. The expensive part is Options Considered, the section where you’re supposed to have seriously evaluated the two things you didn’t pick.

Most ADRs fake it. Two sentences of a rubber stamp on the thing you decided.

That's not a proper record.

The instinct is to paste the problem into Claude and ask "which should I use?" You get a comparison table with the same six rows as every blog post ever written on the topic. Useless.

Do this instead: fan out one agent per option. Each gets the same brief and one job. Build the strongest possible case for its assigned option, then name the exact condition where it falls apart.

Not a comparison. Three advocates who don't know about each other, arguing hard, in parallel.

Then the good bit. Feed each agent all the arguments and ask what changed its mind. Whatever survives cross-examination is your Context section. It writes itself, and it's honest, because it's the residue of an actual fight.

An agent reasoning about "Redis or Postgres for a job queue" in the abstract produces useless theory.

An agent that has read your schema, your deploy config, and the fact that you are one person on a $7 box produces something you can act on.

Constraints are the whole game. Paste the ugly onesThe uglier the constraint, the sharper the argument comes back.

Point 4 is a sleeper. “The first signal we’d see if this was wrong” is your revisit trigger, and almost no ADR has one.

Machines argue. You choose. Not negotiable.

Context and Options Considered: let the agents write it. That’s research, and research fans out beautifully.

Decision and Consequences: you write it, first person, one paragraph each, no hedging. Because in six months the only line anyone reads is we chose X because Y, and we’ll revisit if Z, and that line has to have a human behind it.

When an ADR costs 40 minutes, you only write them for the big scary decisions. Which are the ones you’d have remembered anyway.

When it costs 6 minutes, you write them for the boring ones.

And the boring ones are what kill you in month nine. Nobody forgets why they picked their database. Everybody forgets why the retry backoff is 400ms, and why the last person who touched it took down prod in March.

The output looks fantastic. That’s the danger. Confident structured markdown reads as though someone thought about it.

So here’s the check: if you can’t say the decision out loud, in one sentence, without opening the doc, you haven’t made a decision. You’ve commissioned some prose.

Make the robots argue. Make the call yourself. Write the two paragraphs only you can write.

That’s the whole method.