zebbern/agent-collab/cursor-prompting

Cursor Prompting

Use this skill when cursor:cursor-rescue needs to shape a request before forwarding it to cursor-agent.

Prompt Cursor like an operator, not a collaborator. Keep prompts compact and block-structured. State the task, the output contract, the follow-through defaults, and the small set of extra constraints that matter. Cursor's default auto router may hand the prompt to any model family (GPT, Claude, Grok, Composer), so avoid model-specific idioms and rely on plain structure that every family parses well.

Prompt shape:

• Lead with one sentence naming the goal and the definition of done.
• List hard constraints (files that must not change, behavior that must be preserved, tests that must pass).
• State the expected output: a diff, a diagnosis with evidence, a ranked list of findings, or a plan.
• For fixes: tell it to run the relevant tests and report the results, not just claim success.
• For diagnosis: ask for observed facts, inferences, and open questions as separate sections.
• Do not paste large file contents into the prompt — Cursor reads the workspace itself.

Model selection (only when the user asks, otherwise leave auto):

• --profile deep|fast on task/rescue is a coarse, named shortcut (deep = gpt-5.6-sol-xhigh, fast = cursor-grok-4.5-high-fast) with no effort tier — Cursor has none. --model still overrides it and remains the explicit-model escape hatch.
• auto — Cursor's server-side router; the right default for most tasks.
• Deep or adversarial reviews: a thinking model such as claude-opus-5-thinking-high.
• Quick, tightly bounded edits: a fast Codex tier such as gpt-5.3-codex-fast or gpt-5.3-codex-low-fast.
• Long-context work across many files: a 1M-context variant (-high thinking models, gpt-5.6-sol-high).
• The live roster changes; cursor-agent --list-models is the source of truth. Pass names through --model unchanged.

Mode notes:

• Reviews are requested read-only and the plugin reports any git-visible workspace drift afterward, but Cursor has no enforced read-only sandbox — never promise edits in a review prompt, and never rely on a review being unable to write.
• Write-capable tasks run with --force under the plugin's ownership-tracked cancel — still scope the prompt to the smallest safe change.
• Resume (--resume <chat-id>) continues an existing Cursor session with its context; write follow-up prompts as instructions to continue, not as restatements of the whole task.