cross-model-prompting

name	cross-model-prompting
description	Write prompts that work when delegating to a different model family or agent runtime via bridge/handoff. Use whenever composing bridge tasks, worker prompts, AGENTS.md-style guidance, or instructions another model will consume.

Cross-Model Prompting

Prompts tuned for one model often rely on invisible conventions. When another model family or CLI reads the prompt, the usual breakpoints are implicit context, tool assumptions, and calibration.

What Transfers Well

• Explicit constraints: "do not modify tests", "max 40 lines", "read-only".
• Verification commands with expected evidence.
• Structured output contracts: section headers, json schemas, exact report shape.
• File paths and file:line evidence requirements.
• Selected prompt assets, when passed by id and included with the task.

What Breaks Silently

• Tool-specific references: "use Task", "check memory", "run /loop". Name the desired outcome unless the target runtime definitely has that tool.
• Session-only context. The worker did not see the main conversation; restate the minimum project and task context.
• Skill vocabulary. Translate kit words into plain instructions when handing off to runtimes that may not know MaaawKit.
• Calibration differences. Some models act too broadly; tighten scope and non-goals. Some over-explain; cap report length and demand evidence-first bullets.

Prompt Shape

Every cross-model prompt should contain:

1. Context: repo path, project shape, relevant files, constraints.
2. Task: precise scope and explicit non-goals.
3. Prompt asset: selected asset id/path when the worker should follow a bundled contract.
4. Verification: commands or checks that prove success.
5. Output: exact report format and maximum length.
6. Assumptions: tell the worker where to list uncertainty instead of guessing.

If the prompt is under five lines, it is probably under-specified.

Practical Rules

• Prefer prompt assets over reconstructed specialist prompts when a matching asset exists.
• Do not encode insults or vibes. Encode constraints, evidence, and acceptance checks.
• Review the output contract, not the model's tone. Missing required sections mean the prompt failed and should be tightened.
• For bridge jobs, pass prompt assets with --prompt-asset <asset-id> or MCP promptAssetId. For handoffs, record promptAssetId so the receiver knows the intended contract.