Telling Technology · Learning in public
All episodes
EP.110 — CONDUCTOR
A Claude Code workflow, for too many tabs

16 tabs. 13 idle. Which ones need me?

Every Claude Code tab knows its own thread. None of them knows the other fifteen, and I was deciding every next move by hand. So I built a conductor: one job that reads every tab, proposes one next step each, and only pings my phone when the decision is really mine. The scarce resource was never tokens. It was my attention.

Sixteen small terminal tab icons on a dark screen, three glowing as running and thirteen dimmed as idle, all flowing into one short queue of three cards marked with decision labels
⌁ 16 tabs → 1 queue · every 15 min · only real decisions ping
What does a tab actually know about itself?
01 The digest

One record per tab, before any AI looks at it.

The count that started this: 16 tabs open, 3 running, 13 idle, and I could not tell you which of the 13 were waiting on me. The first piece is deliberately dumb. A Python script reads what already exists: the window file my VS Code extension writes, the live session leases, the tail of each transcript, my clear-and-pause handover notes, and the ticket board. Out comes one digest.json with one row per tab: how long it has been idle, how full its context is, the last prompt, the last reply, and one small flag that matters more than the rest: does the last reply end with a question?

claude code — desktop
> i have 16 tabs open. before any AI decides anything, give me one row per tab with the facts
conductor_digest.py reads the window file, leases, transcript tails, handover notes and tix. Per tab: idleSec, ctxPct, lastPrompt, lastReply, endsWithQuestion, open tickets. Writes digest.json. If the extension that writes the window file is not running, it refuses instead of guessing.
A dark code panel showing one JSON row from digest.json with fields tab, idleSec, ctxPct, lastPrompt, lastReply and endsWithQuestion set to true, highlighted in amber
endsWithQuestion: true
Recreated for clarity. Field names are the real ones, the tab names are made up.
02 The proposer

Rules first, the model second, and nothing moves on its own.

Every 15 minutes a headless Claude Code run executes one command, /conduct. Plain rules handle the obvious rows first: an empty tab idle for 30 minutes can close, a tab with a handover note can resume. Only what is left goes to the model, which writes one proposed next step per tab with a reason. Then every proposal gets a trust class: close, resume, continue, open a new worker, and so on. Each class can run automatically or wait for me, and right now every one of them waits. Auto-steps stay off until I switch them on. Opening new worker tabs stays manual through 2026-10-05 no matter what.

claude code — desktop
> propose one next step per tab. but nothing runs until I trust it, and opening new tabs is always the last thing to go automatic
/conduct: stage 1 applies the rules and writes pending.json. Stage 2 asks the model only about the rows the rules could not settle. Merged into proposals.json with a class per step. Auto is off until you flip it on the page, and open stays manual through 2026-10-05.
A table of trust classes: close-empty, clearpause, continue, close-done, budget and drain marked as could-be-auto, open marked manual until 2026-10-05, and a master switch labelled auto steps set to off
auto-queued so far: 0
Recreated for clarity. In the first 33 runs, every step waited for me.
03 Approve, and it types

The step lands in the real tab. When the tab is ready.

Proposals show up on a /conductor page on my dashboard, sorted into Needs you, Next steps and Decided for you, each with a label. I read one, edit it if I want, and press Approve. That appends a line to a queue file. The VS Code extension from Episode 70 follows that file and types the text straight into the right terminal, then presses Enter. No port, no clicking into the tab. But it waits: only when the tab shows idle twice in a row, its last turn is done and quiet for 20 seconds, and I am not in that tab myself. A queued step older than 30 minutes goes stale instead of landing late.

claude code — desktop
> when I approve a step, type it into that exact tab. but never while it is still working, and never while I am typing in it
Approve appends to queue.jsonl (by: jacob, plus your edit if you made one). The extension tails it every 20 s and checks: idle glyph stable for 2 ticks, turn ended, 20 s quiet, not your active tab. Then term.sendText(text) and Enter. Older than 30 min: marked stale, not sent.
A dark dashboard page titled conductor with a tab budget meter, filter chips for Go live, Your hands, Architecture, Path and Quick, and three proposal rows each with a label chip and an Approve button
Needs you · Next steps · Decided for you
Recreated with made-up tab names. The real page lists my real projects.
⚠️

Honest part: it has not driven itself once yet.

The whole thing was built in two days, 29 commits between 2026-09-26 and 2026-09-27, and in the first 33 scheduled runs the log says auto-queued: 0 every time. That is on purpose. A tool that types into live terminals is a real power, so it starts on a short leash: forbidden text is refused, it never types into the tab I am using, stale steps are dropped, and a guard stops it after three conductor steps in a row on the same tab. The auto switch only goes on after a test run proves the round trip. Until then, the conductor proposes and I approve.

04 Filtering, not pinging

Only real decisions reach my phone

A conductor that pings me about all 16 tabs would just be a 17th tab. So every question gets a label on a five-step ladder, Go live, Your hands, Architecture, Path, Quick, and only the top three reach Telegram. If the model rates a question too low, a few text rules can raise it, never lower it. Each ping is keyed on the tab, the label and a hash of the question, so the same question never pings twice. Tap any image to enlarge it and read the exact prompt that drew it.

BUILT IN TWO DAYS, STILL ON A LEASH

The fix for too many tabs was not more automation.

The obvious fix for 16 tabs is to let the AI answer its own questions. I did the opposite. The conductor reads everything, decides the easy things on paper, and spends my attention only where a decision is really mine: going live, needing my hands, changing the architecture. Tokens scale. My review time does not. The system that runs my work now manages its own review queue, and the first thing it learned was to stay quiet.

A bold card: on the left sixteen dim tab icons labelled 16 tabs, on the right one phone with a single coral notification labelled 1 real decision
AIOS commits 7b6f266 → 8743d15 · 29 commits · 144 Python tests · runs every 15 min · 2026-09-26/27
06 Steal this

Steal the conductor pattern

My tabs and projects stay private. The pattern travels: a dumb digest that turns every session into one row of facts; rules first and the model second; a trust class per action that starts manual; a decision ladder that decides what is worth a ping; and delivery that waits until the tab is idle and you are not in it. Everything in this episode is free and open — clone it, run it, make it yours.

skill /conduct: rules, then one LLM step per tab cmd digest: idleSec, ctxPct, endsWithQuestion template trust classes: auto off, open manual last assets decision ladder + ping dedupe key
run it # repo soon. the pattern: digest.json → /conduct → proposals.json → Approve → queue.jsonl → term.sendText()

No GitHub? Comment CONDUCTOR on the post and the bot DMs you the link.

Next episode

My AI sessions started leaving each other mail

The conductor talks to me. Next: how my Claude Code sessions talk to each other. Durable mailboxes across machines, a doorbell that wakes the right session by name, and one Telegram word that merges a pull request.

A darkened teaser of a row of small mailboxes on a dark wall, one flag raised and a thin teal line running to a terminal window
drops soon · follow so you don’t miss it