[xizhuomengcontin]

Add: surface-scoped-system-prompt — one pattern, Context & Memory.

The pattern. A harness ships a separate system prompt per invocation surface — interactive, headless, declared task mode — and selects it at launch, instead of sending one prompt everywhere and asking the model to work out which situation it is in.

Why it is not a repeat of agent-modes-by-model-personality. That pattern makes the model the variable: different models, different modes. This one is the opposite finding, and I have the negative control for it. Seven models across six labs driven through the same MiniMax Code surface produced prompts identical line for line except - Model: <id>; change the surface instead and you get three different documents. The two patterns compose, and I have listed the other one under related.

Both implementations are checkable.

• MiniMax Code ships three Handlebars templates at assets/agents/_v2/*/SYSTEM.md.hbs and picks one per --prompt-mode (tui, coding, work). coding swaps the Deliverable Files section for Media Output; work adds an Artifact Completion Contract. Tool schema identical across all three.
• Claude Code switches on the entrypoint — the request carries cc_entrypoint=cli versus sdk-cli — and both prompt and tool set change with it. The identity line is rewritten from "You are Claude Code, Anthropic's official CLI for Claude" to "You are a Claude agent, built on Anthropic's Claude Agent SDK", and the tool count drops.

Disclosure. The character and tool counts in the Evidence section were measured from wire captures in an archive my team maintains, which is why it appears in authors and is labelled in References as the measurement source rather than an implementation. The primary source is MiniMax's own repository, not ours, and the pattern stands without our archive — the MiniMax templates are readable in their repo and the Claude Code entrypoint header is visible to anyone proxying their own session.

Checks. bun run validate:patterns and bun run validate:patterns:content both pass on this file (the one remaining content error in the tree is pre-existing, in cross-agent-lesson-sharing.md — missing References). bun run build:data run and its output included. I reverted apps/web/src/data/github-stars.json, since that diff was only the live star count refreshing and had nothing to do with this pattern.

🤖 Generated with Claude Code

[vercel]

@xizhuomengcontin is attempting to deploy a commit to the nkkko's projects Team on Vercel.

A member of the Team first needs to authorize it.

[nibzard]

Thanks — merged with a small maintenance pass: rebased onto the current main (the catalog moved a lot since September 30) and regenerated the site data. The pattern itself was untouched.

🤖 Generated with Claude Code