Start uncertain work

Uncertainty alone does not require research, a requirements brief, a plan, or a work item. Begin with the outcome and add structure only when it helps the next action. Unfamiliar terms—work item, delivery authority, durable context—are defined in the glossary.

Frame the outcome

Give the agent enough to act without inventing scope:

• The purpose and the context that shapes it.
• Observable acceptance conditions.
• Authority boundaries—what it may change without asking.
• The human owner for product choices and residual risk.

Deduplicate the retry logic in the sync workers. Acceptance: one shared
implementation and the existing tests pass. Do not change public interfaces
or add dependencies without asking. I own any behavior tradeoffs.

Let the agent resolve questions answered by the repository. Ask for a focused question only when a product decision would change the outcome or an action needs approval.

Having credentials means the agent can perform an action; it does not grant permission. A configured Git remote does not authorize a push, and production access does not authorize a migration or data change. Grant each shared-state action explicitly when needed.

Choose how much state the work needs

Keep work conversational when it is bounded, can finish now, and does not need a repository record for another person or thread.

Create a work item when its status, decisions, or next action must survive this conversation—for later resumption, coordination, handoff, or auditability. Invoke the agent-work skill and let it find or create the work item and choose the smallest useful set of artifacts. Research, requirements, and plans are optional—not stages to perform by default.

Create a new work item for <goal>.

Use skills for the unknowns that matter

• Use research when a technical question needs authoritative sources, comparison, or findings worth preserving. Keep the answer conversational when durability adds no value.
• Use agent-browser when a web workflow needs running-system evidence rather than source inspection alone.
• Let the agent choose a reversible default for details that do not change scope, safety, or acceptance.

Begin with a slice that can prove something

Ask for the smallest safe action that can produce useful evidence.

Implement the smallest slice that proves <behavior>. Report what changed,
what you observed, what remains unverified, and what still needs a decision
or approval.

For the reasoning behind these choices, see Brief the Agent Like a Capable Co-Worker and Your Repo Is the Memory. For installation and skill discovery, use the Quickstart.