A payment bug exposed that you have no idempotency strategy. Someone has to decide whether to handle it in the checkout service or push it to the gateway wrapper, and that decision shapes every integration you add for the next year. The failure mode here is not a wrong decision. It is a decision nobody can reconstruct: made in a chat window that scrolled away, leaving the team to relitigate it every time someone new asks why.

​
Start in Ask mode

You are gathering, not committing. Ask mode cannot create issues, documents, or plans, so you can explore without watching for side effects, and read-only work runs uninterrupted by approvals.

​
Pick the right instrument

Three shipped options, three different jobs: This one is consequential and hard to reverse, so it wants Advisor, with /research run first if the evidence base is thin. Switch the model selector to Genius for the pass that decides. Leave it on Standard for lookups.

​
Ask for the shape of an answer, not the answer

A prompt that produces something you can act on:
@Advisor We need an idempotency strategy for card payments. Compare handling it in the checkout service versus the gateway wrapper. For each: what breaks if we get it wrong, what it costs to build, and what it forces on future integrations. State your assumptions separately from what you verified. Recommend one, and say what would change your mind.
Three parts of that prompt do the work. “What breaks if we get it wrong” surfaces risk instead of feature comparison. “Assumptions separately from verified” forces the epistemic split you need when auditing later. “What would change your mind” gives you the trigger to revisit.

​
Move it out of the chat immediately

Chat is where the thinking happened; it is not where it lives. Ask the agent to write the decision into a document. This is also the only reliable way to produce an artifact: the Artifacts library registers something when an agent creates durable output, not when a conversation is merely substantial. A long, excellent chat answer registers nothing. The write pauses for a tool approval. That is the system working; read the card, confirm it is writing where you expect, and allow it once.

​
Structure the document so it survives

Six months later the reader is someone who was not in the room. Use the block editor (/ on an empty line opens the menu):
Docs keeps no version history and offers no restore. A deletion cascades to every child document and cannot be undone. Duplicate before any edit you are unsure about; this is the one surface where that habit matters.
Open the issue and use Add context to attach the document. The reasoning then sits beside the work it produced, and whoever opens the issue, plans it, or picks it up in a session starts from the decision instead of rediscovering it.

​
What good looks like

  • The decision is one sentence at the top of a document, not buried in chat.
  • Assumptions are visibly separated from verified facts.
  • The rejected option is recorded with its reason.
  • The document is attached to the issue as context.
  • There is a stated trigger for revisiting it.