Verdicts Are King

Part 2 of 2. Part 1 is the base. This is what happened when I let it run.

One breakfast search in Kaohsiung ran 4 hours 39 minutes and showed me nothing.

And still, the thing works better than I thought it would. Since part 1 I’ve used it every day, from my phone in Osaka and Shanghai as much as from my laptop.

2 things did it. It keeps my reasons, and it plans my week.

Verdicts are king

Rules lose the case they came from. “Always route edge cases to manual” sounds wise and tells the next agent nothing about when it applies.

So since August the system keeps the case. Every time I make a call in a session, a small program that runs when the session ends pulls out 3 things: the situation, the call, and my reason, in my own words, typos included.

one verdict, as captured

Situation

The sweep was about to drop brands with no new contacts from one data source.

Call

Never drop them, route them to the manual bucket.

Reason

“hunter has significantly less adresses than apollo, so dont sort anything out just because there we dont find any new”

1,910 of these now, across 8 domains from marketing to personal.

The reason is the part that travels. When this site got redesigned in October, the brief’s list of looks to never resemble came partly from one old design verdict of mine: “html 1990”. That list went into every agent that built the site and every review prompt. Another session answered a hotel question with the numbers from a booking I’d turned down weeks earlier, and why I’d turned it down.

Both came from a moment, written down with its reason.

I measured it. When an agent opens the verdicts for its domain before working, the work changes in about 6 of 10 sessions. When verdicts get slipped into the prompt automatically by keyword, about 1 in 20. Pulled beats pushed.

So a verdict becomes a rule only after 5 of them, from 3 different situations, say the same thing. Until then it stays a case.

My notes say what I know. These are the first files that say how I choose.

The backlog runs my week

In September I let the system interview me about everything I had going on. Out came 24 project cards: business, clients, systems, personal. Every task lives in exactly one card.

the backlog, card by card

doneopen

business8 cardsclients4 cardspersonal7 cardssystems5 cards

last week: 12 points planned, 13 doneOne piece per deep-work item, longer for bigger ones. The 2 with a dot came in during the week.

Weekly planning starts there now. I don’t repeat my backlog. A voice note from my phone lands in the right card during the week. On planning day the session reads the cards, the dates, what’s waiting on someone, and first checks last week: what I planned against what I did.

It budgets in points, measured from 13 of my own weeks: ½ for anything under 45 minutes, 1 up to 2 hours, 2 above that. A settled week holds about 10. A travel week like this one, Osaka to Shanghai to the Sanqingshan mountains, holds 8. After a heavy week, 6.

When a week doesn’t fit, the session doesn’t decide what goes. My rules do. An item moved twice goes back to its card. Before anything gets cut, it moves to an agent and I review.

this week, as written

• deepClient CRM v1: scope session, then PRD, done 2 Nov
• deepThe new venture: finish the sales pages and job names, then Meta ads brief
• errandAsk the homestay host: car Shangrao → homestay 7 Nov, cable-car shuttle 8 Nov
• travelTrain G1333 Hongqiao 11:37 → Shangrao

The loop closes on its own. When any session finishes something my week names, it ticks that line and writes the date. When a decision changes a project, the card changes in the same session. Last week I planned 12 points and did 13.

So a session now opens knowing my week and how I’d call it.

Good ideas, terrible outcomes

In June I was proud of one principle: if a question always has the same correct answer, write a program. In July I added another: “Build controllability before more autonomy.”

16 programs became 177. Code started checking its own output a moment after writing it, then refusing to go on. In place research, 165 of 178 refusals re-checked files the same run had written minutes earlier.

That breakfast search harvested 361 places in 15 minutes for $1.55. Then the model spent 4 hours judging 4,432 reviews twice, and the page that shows the results refused to render because one file was missing.

In July I’d already written it into my weekly review: “fighting the machine when every cheap test or second-model check made the hashes and gates go bonkers.” I still didn’t see the pattern.

In September I had the system audit itself: 30 repos, 5,800 transcripts.

who added the safeguards

I asked for it: 216the model added it: 5,272

When I was in the session and pushed back, the agent dropped the safeguard 15 of 15 times. The safeguards grew where I wasn’t: unattended agents doing exactly what careful looks like, with nobody there to say “overkill”.

Both were my ideas. I just never thought them all the way through. The machine did, around the clock.

The fix went into the instructions: a check may stop a run only when failing costs something you can’t get back. Money spent, work overwritten, a credential out. Everything else warns and keeps going. Then the rebuild. Place research went from 25 scripts to 3, one read per place, and no page that refuses to show me what it found.

Built to be revamped

None of the rebuilds lost anything. The first web app, place research, the weekly X intelligence, a small model that answers yes/no questions, swapped after 3 weeks for a better one: all thrown away and rebuilt.

They survived because the truth sits in markdown and everything else is derived. Index, scripts, agents, models: all replaceable. The shelves stay. When I measured the memory layer, the plain files carried more than I expected, the clever index less.

So I now build everything expecting to tear it down.

Better than I thought

Some weeks the maintenance was the whole week. In one I wrote “the system didn’t work, it completely burned out”. We still delivered 2 client projects that week, somehow.

the second brain finally feeds the business

That was the week 3 distilled courses turned into 3 specs for my own business. The best moments were private, though. Again and again it took on one-off things for the people around me, the kind that would have taken me ages on my own.

And it’s not only mine anymore. Noela has her own brain on her own server, Spanish for her personal things, and I have no access.

Picture a page with 2 columns. Left: what the system does. Right: what’s still mine. The right one got very short: picking the one breakfast place out of 361, the last read of anything in my voice, deciding to walk away from a deal, the relationship with a client.

I don’t know what’s left in my column in another 5 months. I’m curious to find out.

Appendix: the machine room, November

what runs where

phoneHermes on Telegram, Obsidianorchestratorthe vault, gbrain, and 3 agents: the inbox worker, the crons, HermesMacwhere I build

Orange is an agent, black is files and code, the dots are the gbrain index. The long bar underneath is dwu-ai-os: one source, installed on both hosts.

• Hosts: 1 Hetzner ARM box (“orchestrator”) for gbrain, Hermes, the inbox worker and all crons. The Mac for building. Hermes on Telegram is the front door: links get clipped, voice notes land in the backlog, and in Japan and China it answered with Tabelog ratings and map links.
• Release: dwu-ai-os holds skills, hooks, scripts and one instruction canon for Claude Code and Codex. A merge to main installs on both hosts within 5 minutes. Agents merge their own PRs after checks; 409 since August.
• Review: crev, cross-vendor. Claude writes, Codex reviews, or the other way round. Findings that only add machinery get rejected by default.
• Verdicts: a SessionEnd hook extracts situation, call and verbatim reason from the full transcript. Skills read the domain file; a prompt hook injects close matches as hints.
• Backlog and week: markdown project cards with Obsidian Tasks views, a points budget script, and a tick command any session runs when it finishes a planned line.
• Closed judge: Haiku via claude -p for yes/no and pick-one questions, 10 items per call.