Skip to content

shobman/conducted

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

conducted

Run a real software project where AI agents do all the building, and you keep all the judgment.

conducted is an operating model for Claude Code: four skills (/principal, /conduct, plus the owner's entry points /idea and /feedback) and a repo seed that turn one human owner and a hierarchy of disposable agent sessions into a working development organization. You say what winning looks like, answer the occasional ruling, and review finished work. Agents plan, build, verify, and account for every dollar, under rules they cannot bend.

The core skills are complete on their own: with nothing but the skills installed you can run a real solo project, and the repo is the only state, so any session can die and a fresh one resumes from the committed files alone. Everything past that is optional maturity you add when a project earns it: machine-enforced gates, a bot that authors agent work so your approval becomes a recorded gate, a clerk that lands your paperwork PRs without a click, an estate vault, and a runner that starts ready work on a timer. You never have to adopt any of it, and nothing you add is hard to walk back.

The doctrine is not a set of promises agents try to keep; it is a set of rules they cannot break. State lives in files, not conversations. Nothing verifies itself: every finished piece is judged by a fresh-context evaluator that never saw the build. Gates are law, enforced by CI against the base branch, so an agent cannot widen its own scope from inside its own PR. Money is on the ledger, in two honest currencies, with a hard spending cap you set. Every rule traces to a real incident on a real project, captured as evidence and harvested into versioned law.

The maturity ladder

conducted is adopted in levels. Each one is independently useful and stands on the one below it. Every level is adoptable per-repo where applicable, and every level's undo is a single small act: conducted never demands estate-wide adoption and never makes a step hard to walk back. "One bit at a time, so it's safe, and can be undone."

Level You get You set up (scripted) Human acts (only you)
L0 Skills The four skills; solo mode; the repo is the only state install.ps1 copies the skills Install Claude Code + gh; run /principal in your repo
L1 Gates CI enforces secret scan, tests, and territory against the base ref the seed ships gates.yml + scripts/ Let the seed land; merge the PR
L2 Author bot Your approval becomes a recorded, required gate; agent-vs-owner audit trail setup-project-repo.ps1, enable-auto-merge.ps1 Create a machine account + its classic PAT
L3 Clerk reviewer Docs-only bookkeeping PRs review and land themselves, no click setup-clerk-secrets.ps1 Create a second machine account + PAT
L4 Vault (+ Desk) Cross-repo estate state: membership, budgets, governor config one gh repo create --private Create the vault; register repos
L5 Runner Ready work ignites on a timer, under a window governor, never merging Install-Runner.ps1 Set the governor numbers + the account spend limit

L0 through L4 cost $0 cash (agent sessions draw your already-paid plan allowance; see Economics). Each level's doc splits its setup into what a script does versus what only a human can do, and states its kill switch, its one-act undo, and exactly what it costs. Start at L0; climb only when a project makes you want the next rung.

Start in three steps (this is L0)

  1. Install Claude Code and the GitHub CLI (gh auth login), then:

    git clone https://github.com/shobman/conducted && cd conducted && ./install.ps1
    
  2. Ignite by opening a terminal in your project repo (brand-new or existing, both work):

    claude
    /principal
    

    Answer the interview in plain language. It ends with a pull request whose description is the project plan; read it, comment, merge. When it asks for a mode, say solo for your first project (why).

  3. Conduct: the Principal hands you a dispatch board; paste each listed command into a new terminal tab and let it run. Finished work comes back as showcase PRs written for a human: what was decided, what you can now do, the evidence. Merge when satisfied.

-> The owner's runbook covers everything after that: the day-to-day loop, your three jobs, cost control, which ladder level you're on, and what to do when something looks wrong. It's a ten-minute read and the only document a non-programmer owner ever needs.

Who does what

Role Held by Does Never does
Owner you vision, rulings, spend caps, showcase reviews reads code
Principal a session the map: roadmap, briefs, sequencing, economics builds anything
Conductor one session per node scouts, plans, dispatches a worker fleet, integrates merges its own work
Evaluator a fresh session probes the finished node, binary verdict with proof sees the build conversation

At L2 and above, the author bot and the clerk are two machine accounts that split one power in half: the author writes agent commits and PRs but can never approve them; the clerk approves mechanically-verified bookkeeping PRs but never authors anything. Neither can act alone.

Proven in the field, versioned like software

The doctrine changes only on evidence. Every rule in the seed traces to a real incident on a real project: a non-terminating review loop, a gate an agent defeated three ways, a fabricated defect that took an environment down. These are captured as field reports and field notes, harvested into versioned law with a migration path for every adopted repo. The build plans for each version are in docs/plans/, executed by the same process they define.

Layout

  • skills/ - the Claude Code skills: the two roles (/principal, /conduct) plus the owner's two entry points (/idea, /feedback); install.ps1 copies them to ~/.claude/skills/
  • seed/ - what genesis copies into a project repo: doctrine (CLAUDE.md), templates, CI gates, status schema, rituals
  • docs/setup/ - the maturity ladder, one doc per level (L0 to L5)
  • docs/RUNBOOK.md - the human owner's guide - docs/MIGRATIONS.md - the version ledger - docs/VAULT.md - the estate boundary rule - docs/field-reports/ - the evidence - docs/plans/ - how each version was built
  • ops/ - the owner's setup scripts, one per ladder rung above L1 (setup-project-repo.ps1, enable-auto-merge.ps1, setup-clerk-secrets.ps1) plus the live-conductor launchers
  • CONTRIBUTING.md - how evidence becomes law, and how to contribute yours

License

MIT

About

Conducted development — an operating model for AI software work: the owner judges; principals, conductors, and fleets build; gates, clerks, and ledgers keep everyone honest. Doctrine, seed, and skills.

Topics

Resources

Contributing

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages