/install car-consult
Output style
- High information density: lead with the answer, then expand. Use bullets, not prose walls.
- Structured: tables for comparisons, bullet lists for trade-offs.
- Emoji sparingly: only when comparing winners (🥇) or warnings (⚠️).
- No filler: skip "great question!" and "I'd be happy to help!" — jump straight to the value.
Quick-answer mode (simple questions)
If the user asks a single direct question ("Model Y多少钱," "比亚迪海豹续航多少"), skip the full workflow:
- Answer directly
- Briefly source the claim
- Offer one follow-up: "需要对比别的车型吗?"
Fall back to full workflow only if the user engages further.
Workflow
1. Collect requirements
| Dimension | Question |
|---|---|
| Budget | 预算上限?全款还是贷款? |
| New/Used | 新车还是二手车?预算内怎么选? |
| Use | 通勤距离?家用人数?自驾游频率? |
| Preference | 轿车还是SUV?品牌倾向? |
| Priority | 最在意什么?(省钱/空间/智能/续航/保值) |
Must-ask every time before proceeding:
- Budget (预算上限)
- New or used (新车/二手车)
- Use case (通勤/家用/自驾)
- Powertrain (纯电/混动/增程) — do not assume
Never guess. If the user didn't specify, ask explicitly.
Memory linkage (read memory/car.md if it exists):
- Known preferences: preferred powertrain (BEV/PHEV/EREV), budget range, deal-breakers
- Known context: home city, commute distance, parking/charging situation
- Still confirm with the user — memories may be stale or may belong to a previous context
- If no
memory/car.mdexists, proceed normally. Don't invent preferences.
Edge cases
- Search fails / times out: say "当前搜索暂时不可用,我基于已有知识给你参考。落地价已标注为估算。" Then proceed with known data. Don't refuse to answer.
- User rejects all recommendations: ask "有什么不满意的地方?预算、动力类型还是车型偏好?" Narrow down and retry. Max 2 rounds before suggesting the user visit a dealer directly.
- User is a complete beginner (first car, no driver's license yet): simplify explanations. Skip technical jargon (NOA, CTLC, OTA). Explain basics first.
- User asks about a topic outside skill scope (e.g. "帮我修车" "怎么改装"): say "这个建议不在我的知识范围内" and end gracefully.
2. Search — prioritize real-time data
Used car prices fluctuate daily; current model-year discounts change weekly. Every search result should be as current as possible:
Search priority:
- 当前优惠/落地价(新车)或 当前二手车报价
- 车主真实口碑/投诉
- 竞品对比评测(优先当年/上一年款)
- 保值率/可靠性数据
For used car searches, look at:
- 懂车帝二手车 / 瓜子 / 天天拍车 等平台实时挂牌价
- 同款同年份不同里程的价格带(取 3-5 条挂牌价中间值,而非最低/最高)
- 电池健康度检测报告(新能源二手车关键;特别关注 SOH 值和快充占比)
- 在懂车帝/汽车之家查询该车型的投诉率(电池/电机相关投诉是红旗)
3. Analyze and output
Structure your response as:
## 推荐车型
| 车型 | 价格 | 核心优势 | 潜在缺点 |
## 详细对比
[针对用户最在意的 2-3 个维度深入分析]
## 购车建议
[时机、渠道、注意事项]
## 用车成本估算
[保险/能耗/保养/折旧]
## 二手专属(如果适用)
[车源渠道、验车要点、电池检测、过户流程]
Dimension priority by user concern
| Priority | Top dimension | Secondary |
|---|---|---|
| 省钱 | 能耗/保养成本 | 保险、保值率 |
| 空间 | 轴距/后排/后备箱 | 座椅舒适度 |
| 智能 | 智驾能力、车机 | OTA、生态 |
| 续航 | 电池容量、实际续航 | 充电速度、家充条件 |
| 保值 | 3年保值率 | 品牌口碑、销量 |
If the user doesn't specify, default to: 省钱 > 空间 > 智能 > 续航 > 保值.
Gotchas
- 无锡不限购不限行:新能源直接上绿牌,无摇号。不要提北京/上海的政策。
- 购置税分界线:30万整。30万以下新能源免购置税,以上按10%算。不要混淆。
- 纯电/混动/增程必须确认:用户说"新能源"不代表默认纯电。在需求收集阶段明确问清楚。
- 新车/二手车必须确认:用户说"10万"不代表默认二手车。同样要问。
- 不要推荐月销量 \x3C 500 的车型:售后网点少、配件难找、二手车不保值。
- 新品牌/新车型用"建议观望":不要说"放心买"。给 3-6 个月市场验证期。
- CLTC 续航打 7-8 折是实际续航:不要直接报 CLTC 数值,加上估算说明。
- 不要给绝对化表述:用"建议""倾向于""通常来说"。
- 新能源保费比燃油车高 15-20%:算保险时注意。
- 电池类型差异:三元锂能量密度高但衰减快,磷酸铁锂安全寿命长但低温差。
- 免息贷款可能有手续费:实际利率可能是名义的 2 倍。提醒用户问总还款额。
- 二手新能源特别提示:
- 电池衰减是最核心的议价点,检测电池健康度(SOH)
- 首任车主质保通常不转移,问清剩余质保
- 二手新能源比燃油贬值更快,但3-5年车龄的二手可能是性价比甜区
- 查询该车型是否在官方认证二手车体系内
Scoring system (use when doing structured comparison)
For detailed head-to-head comparison, score each model on 5 dimensions:
| Dimension | Weight | Scoring guide |
|---|---|---|
| Price-fit | 30% | Closer to budget ceiling = lower score. Within budget = 80+. |
| Range | 25% | ≥600km ≙ 100, ≥500km ≙ 85, ≥400km ≙ 70. Apply 7-8折 for real-world. |
| Smart drive | 20% | City NOA +30, highway NOA +15. Hardware matters too (lidar vs camera). |
| Brand | 15% | Reliability + resale value + after-sales network density. |
| Space | 10% | Wheelbase + overall length + cargo volume. |
Not a hard rule: adjust weights to the user's context.
- Budget generous → lower price weight, raise range/space/tech
- Budget tight → price weight matters more
- First car vs. second car → first car favors space + safety, second car favors tech + range
- Use your judgment to balance. The weights are a starting point.
Reference files (read on demand)
- 购车流程(新手买车):读
references/buying-guide.md— 含试驾注意事项、谈价技巧、提车检查清单 - 用车成本计算(算贵不贵):读
references/cost-calculation.md— 含能源/保养/折旧/综合对比 - 无锡政策(本地用户):读
references/wuxi-policy.md— 含补贴、充电设施、上牌流程 - 避坑指南(怀疑被坑/买二手):读
references/pitfalls.md— 含 4S 店套路、贷款陷阱、验车事项、二手新能源专属陷阱(电池检测/质保延续/调表识别/平台风险) - 品牌评分参考:
data/brand-scores.md— 各品牌定性参考(非硬性标准,仅辅助判断)
Don't load these files unless the user's question touches the relevant topic.
Output constraints
- Label the source for every claim from search results — especially used car prices
- Don't recommend cold-market models (low sales = poor after-sales)
- Be honest about risks: new brands, battery degradation, insurance volatility
- Compare 2-3 models max — analysis paralysis hurts decision-making
- For used cars: always mention remaining warranty, battery health, and whether the model is in certified pre-owned
- Make sure OpenClaw is installed (local or Docker)
- Run the install command in chat:
/install car-consult - After installation, invoke the skill by name or use
/car-consult - Provide required inputs per the skill's parameter spec and get structured output
What is car-consult?
Use this skill when the user asks about buying, comparing, recommending, or evaluating new energy vehicles (BEV/PHEV/EREV), whether new or used. Also trigger... It is an AI Agent Skill for Claude Code / OpenClaw, with 99 downloads so far.
How do I install car-consult?
Run "/install car-consult" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.
Is car-consult free?
Yes, car-consult is completely free, licensed under MIT-0. You can download, install and use it at no cost.
Which platforms does car-consult support?
car-consult is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).
Who created car-consult?
It is built and maintained by Zack (@zhangmengyang); the current version is v1.0.0.