Confirmation habit
Executes irreversible operations (ship, delete, external action) without checking first.
Explain impact and wait for confirmation before irreversible operations.
Skip to content01 02 03 
Reviewable · Reversible · Observable
Stop correcting thesame Agent behavior.
PD turns repeated corrections into reviewable, reversible behavior principles, so the Agent handles the next similar task your way.
You approve every principle first. Its effect stays observable and can be rolled back.
A concrete behavior change
Not a saved note. A changed next action.
From repeated correction through Owner approval to an observable change in later behavior.
- Before
Repeatedly expanded scope without explaining the impact.
- Detected
The system found three related corrections.
- Proposed
Explain impact, risks, and verification before expanding scope.
- Decided
The Owner edited and approved it.
- Later
On the next similar task, the Agent asked first.
- Control
Roll back if it causes side effects.
Governance loop
The system brings evidence and proposals. The Owner keeps judgment.
- 01Behavior evidence
- 02Principle proposal
- 03Owner review
- 04Reversible activation
- 05Later behavior observation
- Not every mistake becomes a principle.
- PD does not make value judgments for the Owner.
- No principle activates without review.
- Weak evidence can be deferred or archived.
What PD governs
PD governs the behavior gap between you and your Agent.
Not tool errors, not one-off task failures — but the behavior patterns that recur across sessions and are worth distilling into principles.
PD governs
Behavior-pattern · cross-session human–Agent gaps
- Lacks a confirmation habit before irreversible operations.
- Leans aggressive rather than conservative in refactors.
- Expands scope without explaining the impact.
Not PD
Tool / task level · handled by host & session memory
- A failed git push, a missing flag — handled by the host / runtime.
- Malformed output or JSON — handled by the host's retry & repair.
- A repeated same-task failure — handled by session memory.
Where it helps
Not a single error — a recurring character drift.
These are what PD actually governs: behavior patterns that recur across sessions and are worth distilling into principles.
Refactor restraint
Leans toward broad rewrites over minimal changes, introducing extra risk.
State scope, risks, and the verification plan before broad changes.
Impact transparency
Expands scope or makes claims without citing sources or reasoning.
Cite sources for important claims; explain impact before expanding scope.
The compound effect
One correction changes one action. Accumulated principles settle into character.
Every principle you approve lays groundwork for the next similar situation. Corrections do not evaporate — they accumulate, until the Agent acts your way without you spelling it out each time.
- Sediment, not amnesia
Corrections settle into traceable principles, instead of being re-typed in the next session.
- Compound, not one-shot
Each principle lays groundwork for the next similar context; character accrues, it is not configured once.
- Less attention, not more
You do not watch every conversation — PD surfaces only patterns worth your review; the rest is handled quietly.
Install
Start with the next correction.
After installation, you still control review and activation. You currently need Node.js ≥ 18 and a working OpenClaw environment.
Available nowOpenClaw
Ideas & Research
Explore the judgment an Owner cannot outsource
Read PD's ongoing research into Owners, Agent behavior boundaries, and governable collaboration.
The Expanding Boundary: The Tower, the Boat, and the Silicon Amoeba
As Agents rewrite the boundaries of corporations, society, and meaning systems, how do humans as Owners hold onto the non-outsourceable? From ExO to MVO — exploring the Minimum Viable Organization and where PD stands in the intelligent flood.
“If you’re not suffering, you’re not building value.”
Burn Pain, Co-Evolve