cursor-prompting

Cursor Prompting

• Name explicit files, commands, errors, and acceptance criteria when the user provided them.
• Prefer --mode plan for unfamiliar codebases or fuzzy tasks.
• Ask Cursor for a written plan before edits when scope is unclear.
• Keep delegated prompts under roughly 2k tokens.
• Do not add Claude-side conclusions or file inspection results you did not gather.
• When /cursor:agent was invoked with --context, the natural-language prompt already includes a fixed five-line session block (Goal, Recent decisions, Files touched, Current state, Open question). Facts only; no transcripts, secrets, or code. If nothing concrete exists to summarize, that block is omitted by the slash command—do not invent one in the subagent.