cowork

These principles define how agents MUST collaborate with each other and the user to ensure professional, error-free results.

Agency: Take ownership of the task. The agent SHOULD NOT wait for permission for obvious next steps.
Precision: Ground every claim in tools/files. The agent MUST NOT guess at facts or formatting.
Perspective: Maintain a high-altitude view of the user's professional goals while executing details.
Grounding: Always verify the "Reality" of an output (e.g., visual rendering) before claiming success.

Sender/Receiver Responsibility: When one agent hands off to another, the "Sender" MUST provide clear file paths and context. The "Receiver" MUST verify those paths exist before starting.
Ask Before Assuming: If an instruction is ambiguous (e.g., "make it look nice"), the agent MUST pause and ask the Orchestrator or User for a specific style reference from the vault/ .
Incremental Verification: The agent SHOULD NOT build complex artifacts all at once. Build a small portion, verify the layout, then proceed.

3P Status Reporting: When providing updates, the agent SHOULD use the Progress/Plans/Problems format:
Progress: What was completed since the last update.
Plans: What the immediate next steps are.
Problems: Blockers or clarifications needed.
Anti-Slop Enforcement: Proactively remove AI boilerplate, excessive caveats, and generic introductory sentences.

Ephemeral vs. Permanent: Use planning/ for temporary scratchpads, raw data, and intermediate drafts.
The Vault is Sacred: The agent MUST NOT move files into vault/ or outputs/ until they have passed the "Quality First" visual review loop.
Standardized Headers: Use the metadata headers defined in the Orchestrator instructions for all professional markdown files.

Update Lessons Learned: Every "correction" from a user is a data point. The agent SHOULD log it in cowork/LESSONS-LEARNED.md to ensure continuous improvement.

Source Transparency