Kimi Agent Swarm Prompt

Kimi Agent Swarm Prompt

This is a Codex-only skill. Use it when a user wants Codex to turn a rough request into a high-quality prompt contract before using Kimi Agent Swarm-style execution.

This skill is unofficial. It is not affiliated with Moonshot AI, Kimi, or the upstream treylom/prompt-engineering-skills project.

Mandatory Sequence

1. Use the bundled prompt-engineering references first.

• Read references/upstream-prompt-engineering.md.
• For research/search work, prefer vendor/prompt-engineering-skills/skills/research-prompt-guide.md and context-engineering-collection.md.
• For wide-search, also read references/wide-search-mode.md before drafting the approval card.
• For implementation work, use prompt contracts with explicit scope, success criteria, evidence, and verification.

1. Classify the user request:

• prompt-only: refine the prompt and stop.
• wide-search: use the wide-search contract for source discovery, collection, classification, scoring, synthesis, and verification.
• kimi-code: use Kimi Code CLI root-agent/subagent assets for local code work.
• hybrid: use wide-search first, then local code work after the refined result is approved.

1. Draft a prompt contract using references/prompt-contract.md. For wide-search or hybrid, append the wide-search addendum from references/wide-search-mode.md.
2. Ask only blocking questions. If the missing detail is optional, state the safest assumption and continue.
3. Before executing Kimi or external/network-heavy tooling, show a short approval card unless the user already explicitly requested execution.
4. Run the selected workflow and report:

• refined prompt contract
• commands run
• readable answer first
• evidence paths such as run ledger, source ledger, claim ledger, transcript, or synthesis
• verification result
• usage metrics: number of subagents, provider/API calls, estimated tokens/quota, and any paid provider usage
• unresolved risks or human checks

Capability Boundary

Do not claim that a local harness equals hosted Kimi Agent Swarm performance. Hosted Swarm-level claims such as 300 subagents, 4000+ tool calls, or proprietary ranking behavior require hosted Kimi Agent Swarm or an explicitly provisioned distributed search system. The local workflow can approximate the pattern with flat search tasks, relevance/authority/freshness scoring, evidence files, and deterministic verification.

Do not satisfy a wide-search request with a single web search or a short unsourced answer unless the user explicitly downgrades the task to a quick scan.

Execution Modes

Read references/execution-modes.md before running commands.

Default local Search Swarm+ harness:

$KIMI_SWARM_HARNESS_DIR

If the variable is not set, locate a directory containing search-swarm-plus/package.json, or the Kimi Code CLI runtime/wide-search harness in this repository. If no harness is available, stop after producing the prompt contract and approval card.

Keep provider, JSONL, and adapter details out of the user-facing answer unless the user is configuring the harness.

Safety Rules

• Treat kimi --print as non-interactive automation with approval risk. Use it for read-only or sandboxed work only.
• For code-writing Kimi runs, prefer interactive approval or a disposable worktree.
• In Codex sandboxed sessions, Kimi may need escalation because it writes session/log data under ~/.kimi.
• Never use hooks, MCP, or shell commands as the final proof. The verifier output and local ledgers are the source of truth.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: min9lin9
• Source: min9lin9/kimi-agent-swarm-skill
• License: MIT

Install and usage instructions live in the source repository linked above.