Delegation & Effort: 5 Laws for Working With AI

01What you will learn

• The five laws of delegation, and why the fourth — the human in the loop — is the brake.
• How to split labour between a light, a mid-tier and a strong model, and a human.
• How to choose a model and an effort level by the stakes, not by habit.
• How to verify delegated work without trusting the self-report alone.

02Before you start

• Access to Claude (chat or Cowork) 🌐 global model with the menu for choosing a model and effort.
• One recurring task that contains an irreversible action: moving, deleting or sending files.
• Readiness to say “yes” or “no” — the human in the loop is you.

03Steps

1. Why we need laws

   When you hand work to AI helpers, the costliest mistake is not “a bad paragraph” but an irreversible action in the wrong direction. A typical trap: an automated helper tidies a temporary folder and saves its own draft notes, while the only copy of an important file stays there and is wiped at the next clean-up. The system “saves” the junk instead of the treasure. The five laws stop exactly that.

2. The five laws

   Law	What it says
   1 · No one judges above their level	The light model WATCHES (counts, flags, does not touch). The strong one DECIDES where. The mid-tier one MOVES to a given address.
   2 · The executor does not guess, it is given an address	The order contains an exact target folder, not “decide where it belongs”. The executor is an archivist, not a matchmaker.
   3 · Closed list	Move only into places that already exist. No new places are invented.
   4 · Human in the loop — mandatory	The executor reports “I will move A→X, B→Y” and WAITS for an explicit “yes”. No confirmation, no irreversible action.
   5 · Save the treasure, not the junk	Important things leave temporary storage for the permanent folders. Leftovers of the draft process do not enter them.

   ⛔

   Law 4 is not politeness, it is a fuse

   A move is irreversible for the process. So report the intent, list the “source → target” pairs and wait for the explicit “yes”. Silence is not consent.
   Order template for an executorcopy

   Move the files from <input-folder> into EXISTING folders.
   Do not create new folders. Do not guess the topic — the target is given: <list file → folder>.
   First show a “file → folder” list and STOP.
   Move nothing until I write “yes”.

3. Division of labour

   Principle: if a registry remembers for you, your mind stays free for the next thing, not for the filing. The light model keeps the process awake; the strong model and the human think.

   Role	What it does
   Light model · border guard	Catches every new item the moment it appears and writes a row in the registry: identifier, location, date. It does not read meaning — only “it exists, here it is, it is registered”.
   Mid-tier model · doctor	Reads content, fills in kind and status, checks against a standard. Wakes the human only when a review is needed.
   Strong model + human · the brain	They do not maintain the process. They do the next thing. The registry remembers for them.

4. Model and effort by the stakes

   The model menu next to the send button controls three separate things: the model, the effort and thinking. You can change them at any point — a change applies from the next response. On Fable 5.1, Opus 5.5 and Sonnet 5.5 thinking cannot be turned off in the app; a thinking toggle exists on older models (on Haiku 4.5 it is called “Extended”). Higher effort gives a more thorough answer but is slower and uses more tokens, so you reach your usage limit sooner.

   Model (as of 01.10.2026)	For what
   Fable 5.1	Demanding reasoning and long-horizon agentic work; when Opus 5.5 at higher effort still falls short.
   Opus 5.5	The default starting point for most tasks.
   Sonnet 5.5	The best combination of speed and intelligence.
   Haiku 4.5	The fastest; for small, high-volume work. It has no effort setting.

   Default effort according to the official documentation: Fable 5.1 — high, Opus 5.5 — medium, Sonnet 5.5 — high. Thinking is adaptive: always on for Fable 5.1 and Opus 5.5; on Sonnet 5.5 the model decides how much to think.

   Level	For what
   Max (and Extra high for long agentic work)	Anything that goes out or up for approval: offers, contracts, legal texts, pricing and financial models, funding applications, new knowledge others will rely on.
   High (our working default; the default in the menu depends on the model)	Drafts, brainstorming, analysis of status and blockers, standard offers, translation of a finished text.
   Low / Medium	Renaming and tidying, quick look-ups, text extraction, simple formatting.

   ♻️

   AIlyak

   Systematically removing effort until only the value remains. The heavy model is not wasted on routine — but a light setting is not used for anything that leaves the building.

5. Verification and practice

   When a strong model checks work, rely on it to flag the doubtful spots itself — but always ask explicitly “what worries you here?”. Silence is not a clean pass. And never trust the self-report alone: look at the real output.

   🛠️

   Exercise

   Delegate one routine task to a light sub-agent and check the real result, not its report. Any irreversible action along the way happens only with your “yes”. Do, don't consume.

04Check

05What's next

06Sources

1. Models overview — the current model line-up, default effort and thinking (checked 01.10.2026).
2. Change the model, effort, and thinking settings — the effort levels and thinking.
3. KAGAMI working practice, generalised for publication.