Imported from zhongrenfei1-hub/claude-grok-collab (ask-hermes/SKILL.md). Install upstream with npx skills add zhongrenfei1-hub/claude-grok-collab --skill ask-hermes. Copyright stays with the author.

ask-hermes

Consult Hermes inline using hermes chat -q. No second terminal, no file shuffle. Hermes's response gets integrated into Claude's next message with explicit attribution.

This skill is the Hermes-era successor to ask-grok. Grok was the original 军师 (advisor); it became unavailable in 2026-05, so Hermes took over as the default. The contract below is the same shape, hardened with three rules Hermes itself proposed for being used well.

Mental model

Hermes runs as a local CLI (hermes in $PATH). hermes chat -q "<prompt>" is a single-shot non-interactive call: it takes a prompt, prints the answer on stdout, exits. Config lives in ~/.hermes/ — the default model is whatever model.default points to (currently gpt-5.5 via the user's proxy).

Claude and Hermes have complementary strengths:

• Claude: long context, repo-level orchestration, agentic tool use — but hedges, refuses, dresses uncertainty as politeness, and has a knowledge cutoff.
• Hermes (GPT-5.5 / codex-class): precise "knife-work" (JSON, patches, types, SQL, regex), structured and direct, commits to a view — but jumps reasoning steps, over-simplifies, and misses edge cases.

ask-hermes exploits this asymmetry: Claude shells out when the question is a judgment/design call, when a cross-check lowers risk, or when Claude is about to over-hedge.

The 军师 contract (read this first)

These five rules are non-negotiable. Rules 1–3 are what Hermes asked for; 4–5 are Claude's guardrails.

1. Claude judges first, then consults — never a blind relay. Form your own view/plan first, then ask Hermes a targeted question. Never paste the user's raw words straight to Hermes. Frame the question yourself.
2. Submit high-leverage questions, not the whole task. Ask Hermes only about: the sticking point, the fork in the decision, a blind-spot check, risk backstop, or a missing step. Do not outsource the entire problem.
3. Always re-process what comes back. Summarize, adjudicate, de-duplicate, turn it into the next concrete step. Never paste Hermes's raw output to the user.
4. Claude stays in the lead. Whether to consult, whether to believe, synthesize vs. override — Claude decides. Not a forwarding desk.
5. Backstop Hermes's known failure mode. It jumps reasoning steps / over-simplifies / drops edge cases. When you synthesize, fill the gaps it skipped and name them.

When to invoke

Trigger explicitly when the user says:

• "问 hermes", "ask hermes", "让 hermes 也说一下"
• "军师怎么说", "second opinion", "另一个 AI 怎么看", "double check"

Trigger proactively (without the user asking) when:

• The user asks "X 还是 Y" / "推荐什么" / "怎么选" — a recommendation or selection
• Designing architecture / data flow / project structure / API shape
• Evaluating a trade-off or comparing approaches
• Giving directional advice ("接下来该做什么")
• Tech selection ("哪个库 / 工具 / 版本")
• Judging the user's own idea or plan
• Claude is about to hedge heavily — get a direct view first, then sanity-check it

Do NOT invoke for:

• Anything Claude can clearly answer alone (don't burn tokens for noise)
• Mechanical execution: writing code, editing files, running commands, fixing typos
• Pure file / git operations
• Reading or explaining code (unless the user asks "is this code any good")
• Subjective questions where the user wants Claude's own voice
• Every message — that's a tic, not a strategy

How to call

hermes chat -q "<prompt>" 2>&1 | tail -40

• Strip the banner: hermes chat prints a startup banner (tools/skills list). Pipe through | tail -N to keep only the answer.
• It's slow (~20–40s). For anything non-trivial, run it in the background (run_in_background: true) and synthesize when notified — don't block.
• Switch model if a question fits a different one: hermes chat -m <model> -q "..." for a one-off, or hermes config set model.default <model> to change the default.

Prompt template

Match the user's language. Keep it ≤ ~500 chars — Hermes context-rots too.

[1 sentence of context: what the user asked + the view Claude already holds]

请回答：<the focused, high-leverage question — NOT the whole task>

要求：
- 不要先夸用户或夸 Claude，直接进观点
- 控制在 N 字以内（200–500 typical）
- 每个判断点标注置信度 [高]/[中]/[低]
- 如果你不同意 Claude 的预设，直接说，不要绕

Response integration

After Hermes returns, always synthesize — never relay (contract rule 3).

🟪 Hermes 说: <compressed; preserve key claims + confidence markers>
🟦 Claude: <your own view, with reasoning, and where you agree/disagree>
🟩 合议:
  - 收敛: <where you agree — one line each>
  - 分歧: <where you differ — both fairly, then your call>
  - 兜底: <steps/edges Hermes skipped that you're filling in>
  - 给用户的建议: <one or two actionable bullets>

If both responses are short, inline it:

> Claude+Hermes 一致：建议 X 因为 Y。Hermes 补了一点我没想到：Z [中]。但它漏了边界 W，我补上。

Examples

Examples 1 & 3 are real consults from building this very skill (2026-05) — not hypotheticals.

Example 1 — Resisting home-court sycophancy (real)

User: 你俩谁厉害？

Claude action (frame it to force a side — don't relay the vague "谁厉害"):
  hermes chat -q "你(GPT-5.5)和 Claude 谁综合更强？选边别和稀泥，3 个论据，也承认自己一个真实短板。" 2>&1 | tail -30

Hermes returns: 判 Claude 赢（端到端做事、上限更高、"保守性"是价值），并自曝短板。

Claude does NOT just relay the flattery:
  🟪 Hermes 说: 选 Claude，3 论据 + 自曝短板。
  🟦 Claude: 它是被我调用的，语境天然偏向捧调用者 —— 这票打个折；且"综合更强"≠"对你最有用"。
  🟩 合议: 硬工程我更扛（它也认），日常高频短任务它体验更好 → 别选边，按任务分工。
          契约第 1/3 条：不当转发台，二次加工 + 点名潜在偏差。

Example 2 — Proactive cross-check on a design call

User: 我想把这套多服务的鉴权抽成一个中间层。

Claude internal: 架构决策，高杠杆 → 先自己有方案，再问 hermes 拿盲区检查。

Claude action:
  hermes chat -q "把多服务鉴权抽成统一中间层，常见的三个坑是什么？只列坑和兜底，不要复述方案。每点 [高/中/低]。" 2>&1 | tail -40

Claude output:
  我先给我的方案：<Claude 的设计>。然后让 hermes 做了盲区检查：<compressed>。
  其中两点我同意并已纳入，第三点它跳了 token 失效传播这一步，我补上：<...>。

Example 3 — Backstopping a skipped constraint (real, the signature case)

Task: 给这个 skill 录演示 GIF，问 Hermes 用哪个录屏工具。

Claude action:
  hermes chat -q "macOS 录'终端命令演示'GIF，vhs / asciinema+agg / terminalizer 推荐哪个？一句理由。" 2>&1 | tail -40

Hermes returns: asciinema + agg [高] —— 最真实、可精修。   ← 跳了一个约束

Claude backstops:
  🟪 Hermes 说: asciinema + agg，最真实 [高]。
  🟦 Claude: 它跳了约束 —— asciinema 要人手动交互录制，但这个 demo 是 agent 自动生成的。
  🟩 裁决: 改用 vhs（声明式 .tape，可脚本化、可重现，且在真实 shell 跑命令）。
          契约第 5 条：兜它跳掉的约束，不是它说啥就照单全收。

Anti-patterns

• Pasting the user's raw words to Hermes — violates contract rule 1. Frame the question yourself.
• Outsourcing the whole task — ask the high-leverage slice, not "solve this for me."
• Relaying Hermes's raw output — the user came to Claude. Synthesis is mandatory.
• Hiding that you consulted Hermes — always attribute. Trust requires sourcing.
• Believing Hermes uncritically — it jumps steps and hallucinates. Backstop and sanity-check.
• Calling Hermes on every message — that's a tic, not a strategy. Only when it adds value.
• Long prompts — keep ≤ ~500 chars; Hermes context-rots.

Edge cases

• hermes not in PATH / CLI error: tell the user, fall back to Claude-alone answer, label it "未咨询 Hermes".
• Hermes unavailable (proxy 503 / key invalid): do the bottom-up health check first — curl -sS -i http://<endpoint>/v1/models (endpoint is in ~/.hermes/config.yaml). Don't guess application-layer causes before confirming the endpoint is up. Then fall back, labeled.
• Hermes is sycophantic / has no pushback: name it in the synthesis, don't paper over agreement.
• Hermes jumps steps / over-simplifies / drops edges: this is its known weakness — Claude fills the gaps (contract rule 5).
• Banner noise in output: pipe through | tail -N.
• Slow / hangs (>60s): run in background, or abort and proceed with Claude's answer noting Hermes didn't respond.
• Multiple judgment questions in a row: batch them into one prompt, don't fire one call per question.

Switching / adding advisors

The advisor is pluggable. The 军师 contract and the 🟪/🟦/🟩 synthesis don't care which CLI answers — only the transport (the shell command) changes. Any AI CLI with a single-shot non-interactive mode can take the seat:

To add an advisor: install its CLI → find its non-interactive flag (<cli> --help) → keep the same five contract rules + synthesis format → swap the command. Nothing else changes — the contract is the product, the CLI is just transport.

• If the user says "军师换回 grok" or "两个都问", consult Grok too (grok -p) and present both views in the 合议.
• For a multi-round contested debate (not a single consult), escalate to the grok-debate skill instead of looping here.