Attention Deficit AI - Ep 005 - screen artifact 02

AI Operating System

Updated: August 15, 2026
Work-log control plane and usage hotspots
Original work-log system architecture map posted on X
July 2026

Then

Everything worked, but the system carried too much context through long-running sessions.

August 2026

Now

The operating system is a routing layer: scope the context, script the repeat work, and send judgment to the right model.

Signals

Manual intake

Ideas, todos, emails, docs, and experiments entered the same work stream.

  • mail
  • calendar
  • notes
  • drive

Scoped intake

Inputs stay attached to the job they came from before the model sees them.

  • email
  • calendar
  • browser
  • terminal
Router

Master prompt

Standing rules, tools, project context, and session history rode together.

  • AGENTS
  • skills
  • schemas
  • history

Routing contract

Always-on rules stay small; detailed filing, skill, and project rules load on trigger.

  • AGENTS
  • guides
  • skills
  • project scope
State

Broad memory

The work log held the truth, but agents often re-read and rediscovered it.

  • weekly
  • projects
  • content
  • lexicon

Canonical ledgers

Facts, queues, outputs, and project state each have a home and can be read narrowly.

  • weekly
  • INDEX
  • PIPELINE
  • terms
Work

Model-heavy loop

Planning, editing, checking, mining, and cleanup could all become model calls.

  • Opus
  • subagents
  • long chat
  • fix turns

Scripts before models

Consistency checks, session mining, index generation, and content packing run deterministically.

  • check
  • mine
  • pack
  • render
Judgment

Strong model default

Hard days pulled many tasks onto the strongest model, including execution work.

  • plan
  • edit
  • search
  • summarize

Model routing

Planning, edits, bulk work, and final synthesis can travel different model paths.

  • plan
  • execute
  • bulk
  • synthesize
Outputs

Useful artifacts

Posts, status, demos, and screen artifacts shipped, but the path was expensive.

  • posts
  • demos
  • HTML
  • status

Same outputs, cleaner path

The control plane keeps the output surface while narrowing what each task carries.

  • podcast
  • blog
  • briefs
  • reports

Usage Hotspots

work-log
6.6B
AI architect
3.8B
my closet
2.8B
dual view
1.8B
adaptive UI
389M

How It Works

RouteDecide which model, tool, and files the task actually needs.
ScriptMove repeatable checks and assembly out of model calls.
FileKeep durable facts in canonical homes so they are not rediscovered.
MeasureUse session mining to see where work actually gets expensive.