The agent ecosystem course
  • 0Before you start
  • 1Claude Code
  • 2The mental model
  • 3Map your ecosystem
  • 4The brief
  • 5The connectors
  • 6The runbooks
  • 7Working hours
  • 8The ecosystem
  • 9The backup
  • 10Sharing skills
  • 11Getting found

The stages

  • 0Before you start
  • 1Claude Code
  • 2The mental model
  • 3Map your ecosystem
  • 4The brief
  • 5The connectors
  • 6The runbooks
    • 6.01The runbooks, how you do each job
    • 6.02Where they live
    • 6.03Skill or brief, which goes where
    • 6.04How it finds the right one
    • 6.05An example skill
    • 6.06Which jobs to hand over, the five questions
    • 6.07Write your first one, by interview
    • 6.08Test it before you trust it
    • 6.09Ready when, what goes wrong, and homework
    • 6.10[BONUS]One skill handing to another
    • 6.11[BONUS]The LLM Council, five advisors on one decision
    • 6.12[BONUS]The runbook that improves itself
    • 6.13[BONUS]A skill that plays your client
    • 6.14[BONUS]A deeper version, not a bigger one
  • 7Working hours
  • 8The ecosystem
  • 9The backup
  • 10Sharing skills
  • 11Getting found
BACK TO THE COURSE

Stage six, session two

The runbooks

It knows who it is and it can reach your things. Now it needs to know how the repetitive jobs are done here, written down once as a skill.

6.11[BONUS]

The LLM Council, five advisors on one decision

Ask one chat a question and you get one answer, and nothing to check it against.

The council is a skill that puts the same question in front of five advisors at once, each arguing from a different angle and told not to be balanced.

Then all five mark each other's answers with the names stripped off, and a chairman writes the verdict.

It is adapted from Andrej Karpathy's LLM Council, and it is for a decision with money or months attached, not for a factual question or a writing job.

you
One question, with your numbers and what happens if it goes wrong
  • Contrarian

    looks for the fatal flaw

  • First principles

    asks what you are really solving

  • Expansionist

    looks for the upside

  • Outsider

    sees it fresh

  • Executor

    asks whether it can be done

Five answers, each from its own angle and told not to be balanced

  • Contrarian

    marks it

  • First principles

    its answer, out

  • Expansionist

    marks it

  • Outsider

    marks it

  • Executor

    marks it

Every answer to the other four, with the names stripped off

it
The chairman writes the verdict, where they agree, where they clash, one recommendation and one thing to do first
you
You read it, and the decision is still yours

It is a skill, so it installs the way any skill does.

Paste the prompt below into a chat in your folder and it fetches the file from GitHub and puts it under .claude/skills.

Then start a message with “council this” and the question, with your numbers and constraints in it, and read what comes back.

It is at its most useful on the days it tells you something you were hoping not to hear.

Install the council
Please install this Claude skill for me. The SKILL.md file lives in this GitHub repo: https://github.com/oliwoodman/llm-council-skill Set it up so I can start using it, and walk me through anything you need from me.

Note: If you have already decided and want agreement, it will mostly tell you the other thing, which is the point.

The agent ecosystem course

The LLM Council, five advisors on one decision

This is part 6.11 of the course. Start it free and it opens straight away, with no password.

A few quick questions come next. You will also get my Thursday newsletter, and you can unsubscribe from it without losing the course. How your details are kept is in the privacy policy.

Already have access? .