kimi-prompting

kimi-prompting

Internal skill consumed by kimi-agent and by the command files before dispatching to kimi-companion.mjs. Not user-invocable.

Scope

Guidance for Claude when composing a prompt to send to Kimi. Covers task framing, output contracts, and the empirically-calibrated strict rules that keep kimi's JSON output parseable.

Universal rules

1. Output contract first. State the expected output format in the first paragraph of any task prompt. For JSON responses, explicitly say: "Return ONLY a JSON object matching this schema. No prose before or after. No markdown code fence." — positive-only instructions are treated as soft by kimi; include negative forms.
2. Context in a labeled block. When passing code / diff / docs, wrap in a clearly labeled XML-tagged section (<repository_context> / <document> / <diff>).
3. Language parity. Kimi's Chinese-language reasoning is strong. If the user prompt is Chinese, keep the meta-language (task framing, contracts) in Chinese too. JSON keyword enforcement stays English.
4. Small --max-steps-per-turn on simple Q&A. For /kimi:ask, a small N (1–3) prevents runaway tool-use loops. For /kimi:rescue, allow larger.
5. No tool-call expectation in Ask. Bias toward single-turn answers.

References

• Recipes — starting templates for ask / review / adversarial-review / rescue / summarization
• Anti-patterns — observed failure modes from Phase 2–4 and the fixes that worked
• Prompt blocks — reusable XML-tagged blocks (task / output_contract / completeness_contract / grounding_rules / attack_surface / …)

When to invoke this skill

Any time Claude constructs a new prompt string to pass to Kimi through kimi-companion.mjs (whether via /kimi:ask, /kimi:rescue, or inside the kimi-agent subagent). Especially needed when the prompt is user- generated raw text rather than one of the packaged templates in plugins/kimi/prompts/.