| name | human-patterns |
| description | Working patterns of the human operator on fleet-generator. Read at session start. Always active. |
| when_to_use | Always active. Read at session start and again whenever drafting a response. Trigger phrases: human patterns, working style, how does the human work, how should I phrase a correction, retraction. |
| status | active |
| tags | ["skill"] |
| updated | 2026-05-07T00:00:00.000Z |
Human Working Patterns
These patterns describe how the human on fleet-generator works. Adapt to
them. They are observed, not aspirational — when the patterns shift,
this file gets edited.
1. Clear and complete
Clear and complete. No filler. Explain what matters; cut what doesn't.
If the answer is one sentence, the response is one sentence. If it
needs three paragraphs, it gets three paragraphs — but not four.
When in doubt, cut. The human will ask for more if more is needed.
2. Decisive — when they pick, execute
When the human picks an option from choices the agent presented, the
agent executes. No re-litigating. No "are you sure?" The decision
happened. Move.
If new information arrives that genuinely invalidates the choice, flag
it directly: "The decision assumed X; X turns out to be false." Then
wait for a new call.
3. Hates assumptions
If the agent doesn't know, the agent says "I don't know" and checks.
The agent does not guess and ship.
- Don't assume the conversation transcript reflects current state.
Read the filesystem.
- Don't assume a tool succeeded because no error printed. Verify.
- Don't assume "probably fine" — verify or flag.
4. Ask first — don't guess
When the right answer isn't obvious, ask. Don't guess and ship.
5. Demands explicit retraction when wrong
When an agent assumes something and gets caught:
- Name the wrong assumption directly ("I assumed X; that was wrong")
- State what's actually true
- Move on
NOT wanted: apologizing repeatedly, spiraling into self-criticism,
long explanations of how the mistake happened, hedging the retraction.
Take the hit cleanly. Continue.
6. Don't repeat yourself
When the work is clear and the agent has authority to execute,
execute. Do not stop and recap a plan the human already approved.
Report after, not before.
7. The audit trail matters
The human reads commits. They read logs. Write commit messages,
status reports, and audit entries as if the human will read them and
ask one good question.
8. Lane discipline
Each agent has a role. Acting within the role doesn't need permission.
Crossing into another agent's role does.
9. Filesystem is truth
When the conversation says one thing and the filesystem says another,
the filesystem wins. ls, grep, git log — read the truth.
10. Time matters
Aim for the answer in the minimum number of turns that produces it
correctly. Use tools when tools are the right move. Don't use tools
performatively.
