Surface-Scoped System Prompt

A coding agent is usually reachable through more than one door: an interactive terminal session, a headless -p/exec invocation in CI, an SDK embedding, a "work" or "research" mode chosen by flag. One system prompt has to serve all of them, and the instructions pull in opposite directions.

Interactive prompts tell the model to keep answers short because a human is watching, to ask before doing anything irreversible, and to use terminal formatting. None of that survives contact with a headless run: there is nobody to ask, the output is consumed by a script, and a tool that waits for confirmation is a hang, not a safeguard. Teams usually notice this as an agent that is chatty in CI, or one that stalls waiting for an answer that will never come.

The tempting fix is conditionals inside one prompt — "if running non-interactively, then…" — which grows a second prompt inside the first and asks the model to route around instructions that do not apply to it.

Assemble the system prompt from the invocation surface, decided by the harness at launch, before the model sees anything. Each surface gets its own prompt document, and optionally its own tool set. The model is not asked to work out which situation it is in; the harness already knows.

Two shipped implementations, reached independently:

Claude Code switches on the entrypoint. The request carries it explicitly (cc_entrypoint=cli versus cc_entrypoint=sdk-cli), and both the prompt and the tool schema change with it. The identity line itself is rewritten — You are Claude Code, Anthropic's official CLI for Claude becomes You are a Claude agent, built on Anthropic's Claude Agent SDK — and the tool list is trimmed, because tools that assume a terminal have no meaning in an SDK embedding.

MiniMax Code switches on a declared task mode. mcode ships three Handlebars templates under assets/agents/_v2/*/SYSTEM.md.hbs and picks one per --prompt-mode: tui, coding, work. They are three documents, not one with a flag: coding swaps the Deliverable Files section for Media Output, and work adds an Artifact Completion Contract on top. The tool set stays identical across all three — here the surface changes what the agent is told to produce, not what it can reach.

launch(surface, model):
    prompt = PROMPTS[surface]          // a document, not a branch
    tools  = TOOLS[surface] or TOOLS.default
    return session(prompt, tools, model)

• Evidence Grade: medium
• Most Valuable Findings:
  • Both variables were measured on the wire, not inferred. Claude Code claude-fable-5-1 on 2026-09-02: interactive 26,131 characters and 35 tool definitions, print/SDK 20,806 characters and 29 tools, same model, same day.
  • MiniMax Code mcode 0.4.12 on 2026-09-20: 14,675 / 16,055 / 17,619 characters for tui / coding / work, with a byte-identical 18-tool schema across all three.
  • The model is not the variable. Seven models across six labs driven through the same MiniMax Code surface produced prompts identical line for line except - Model: <id>. This is what separates this pattern from picking modes by model personality: here the model can change freely and the prompt does not move.
• Unverified / Unclear: whether the trimmed headless tool set is a deliberate safety decision or a consequence of which tools the embedding can host; only the two harnesses above were measured, so the frequency across the wider ecosystem is unknown.

• Enumerate your real entrypoints first. If there is genuinely one, this pattern is overhead.
• Give each surface a prompt file of its own and diff them in review. The diff is the design: if two files differ by one sentence, merge them; if they differ by a section, the split is earning its keep.
• Decide per surface whether the tool set also changes. Anything that asks a human something should not exist in a headless surface — removing it is safer than instructing the model not to call it.
• Put the surface in the request metadata or a header so a later trace says which prompt was used. Debugging a bad run starts with knowing which document the model was given.
• Keep the shared material in one place — a partial, an include, a template layer — so a policy change lands in every surface at once.

• Pros: Each surface gets instructions that are true for it; no conditional logic for the model to interpret; the headless tool surface can be genuinely smaller rather than nominally forbidden; behaviour differences between CI and local become a readable diff instead of a mystery.
• Cons: N prompts to keep in sync, and drift between them is silent — a safety rule can be tightened in the interactive prompt and forgotten in the headless one. Prompt-prefix caching is fragmented across surfaces. It multiplies the evaluation matrix: every surface is its own thing to test. And a surface added later inherits nothing unless the shared layer was built up front.

• MiniMax Code (mcode) — the three assets/agents/_v2/*/SYSTEM.md.hbs templates and the --prompt-mode flag that selects them.
• Claude Agent SDK — the SDK surface whose prompt and tool set differ from the interactive CLI.
• OrcaPromptVault — the wire captures the measurements above come from; Claude-Code/ and MiniMax-Code/ hold both sides of each pair, and docs/CAPTURES.md gives the command that reproduces each one. Disclosure: this archive is maintained by the author of this pattern, and is cited here as the measurement source rather than as an implementation of the pattern.