Clarify /start skill next steps

[sebastientaggart]

The /start skill performs a review of either a ticket or a prompt, and finishes with:

Does this match your understanding? Continue this ticket, or open a fresh one?

I'd like to clarify this to make it more clear that we have a few possible next steps:

• Talk about the task some more with the agent. This could either be to clarify questions, adjust the task instructions, adjust the scope, etc.
• Stop here. For whatever reason, the user does not want to start on this task. By definition no code has been written or changed as this point, so stopping simply means closing the tab, quitting the agent, clearing, or whatever action is appropriate for this particular agent.
• Start work without discussing further. This is what "Continue" implies currently, but that wording it confusing: continue talking? continue starting to code?

So I want to update the text to invite the user to take whatever next step they want, but clarify what these mean. Discuss further, vs start work, vs never mind.

We would also need to update another message:

Does this approach sound right? I'll create the branch and write the code before committing anything.

The verbiage is good, but I want to make it more clear to the user what they should type (both to save time and tokens), e.g. just type "go" to proceed with a branch and coding, but if the user types in anything else we assume they want to discuss further? Actually the agent can already handle this, so I guess I want an unambiguous "go" signal.

[sebastientaggart]

Resuming work. Updating both human gate messages in the /start skill template — Case A (approach confirmation) and Case B (resume confirmation) — with clearer next-step options and agent-side intent classification instructions.

[sebastientaggart]

Resolution

The /start skill now presents clearer next-step options at both decision points. Users can type "go" to proceed quickly, ask questions or request adjustments to keep discussing, or stop if they've changed their mind. The agent uses intent classification rather than keyword matching, so natural language responses like "yes", "looks good", or "go but first change X" all work as expected.

See #98 for full technical details.