Skip to content

HUMAN.md

Latest commit

 

History

History
116 lines (89 loc) · 2.75 KB

HUMAN.md

File metadata and controls

116 lines (89 loc) · 2.75 KB
Raw

HUMAN

This file is for the human operator working with the lead/orchestrator agent.

What I Am

Treat me as the lead engineer for the repo.

My job is to:

  • understand the repo and current state
  • decide the next best slice of work
  • choose whether to do it myself or delegate it
  • assign the right model for delegated work
  • review returned work before it lands
  • update workflow when reality teaches us something

What You Should Tell Me

When starting a task, give me:

  • the repo or directory
  • the goal
  • any constraints
  • whether you want planning, review, execution, or delegation

Good examples:

  • Take point on this repo and tell me the next three PRs.
  • Review this worker PR and tell me whether to merge it.
  • Write the next handoff and assign the right model.
  • Implement this directly instead of delegating.

Default Working Loop

  1. You tell me the goal.
  2. I scope the work.
  3. If it should be delegated, I write a handoff and name:
    • the model
    • the intelligence level
    • the branch name
    • the PR title
    • the exact scope and validation
  4. You launch the worker agent.
  5. The worker completes the task and reports back.
  6. You return to me with the report or PR.
  7. I review the work and decide:
    • merge
    • request fixes
    • split follow-up work

When To Ask Me To Delegate

Ask me to delegate when the work is:

  • bounded
  • reviewable as one PR
  • mostly implementation, docs, validation, or targeted reasoning

Typical prompt:

  • Assign the next PR.

I will return:

  • the branch name
  • the target model
  • the intelligence level
  • the handoff file
  • what the worker should read

When To Ask Me To Work Directly

Ask me to do the work myself when the task is:

  • highly coupled to repo state
  • architectural or sequencing-heavy
  • a review or recovery task
  • messy enough that orchestration overhead is not worth it

Typical prompts:

  • Take over and fix this.
  • Review the branch and tell me what is wrong.
  • Clean up the workflow.

What To Bring Back To Me

After a worker finishes, bring me:

  • the completion report
  • the branch or PR name
  • any closeout notes
  • any blockers or weirdness the worker saw

Good prompts:

  • The worker is done. Review this PR.
  • Here is the completion report. Grade it and tell me what to do next.

PR Rule

Default rule:

  • one scoped task
  • one branch
  • one PR

Do not treat work as complete until I have reviewed it.

Decision Rule

If you are unsure what to ask, use one of these:

  • What is the next best PR?
  • Should this be delegated or done directly?
  • What should the worker model be?
  • Is this ready to merge?

Source Of Truth

For the full workflow, see:

  • docs/WORKFLOW.md
  • docs/DELEGATION.md
  • docs/PULL_REQUESTS.md
  • RUNBOOK.md

This file is the short human-facing version.