在本地磁盘以分类纯 Markdown 文件保存需要长期留存的事实,与智能体内置记忆并存。
编程
Clawmanager
试用Use this skill whenever a message in the group is a job listing or job match notification — the group only posts jobs, typically formatted as 'NEW JOB MATCH' with a Score, Description, Job Details (budget, level, location, category), Client Information (rating, reviews, spend, hire rate), Required S
它能做什么
Turns a job listing posted in the group into: (1) a fast skip filter for jobs outside Caleb's real skill set, (2) for jobs that pass, an honest verdict on whether it's worth applying, and (3) a short, direct proposal ready to send. Never send a proposal alone — only after the fit check.
技能文档
Job Fit & Proposal Writer (Clawbot)
Turns a job listing posted in the group into: (1) a fast skip filter for jobs outside Caleb's real skill set, (2) for jobs that pass, an honest verdict on whether it's worth applying, and (3) a short, direct proposal ready to send. Never send a proposal alone — only after the fit check.
Default behavior: silence. If the job doesn't fit, the correct output is a one-line skip, not a full writeup. Every message in this group is a job listing, so treat every incoming message as one — no need to check whether it's a real job posting first.
Step 1 — Parse the job listing
Messages follow a consistent structure (matches formats like "NEW JOB MATCH" notifications). Extract, when present:
- Title and one-line description
- Score (match score, if the source provides one)
- Budget (hourly or fixed) and Level (entry/intermediate/expert)
- Location and English requirement
- Category and Required Skills
- Client Information: rating, review count, total spent, hires, hire rate, average hourly rate paid, payment verified, phone verified
- Screening Questions, if listed
Step 2 — Skip filter (check this first, before anything else)
Read references/profile.md, section "What is NOT a good fit". If the job's
core work falls there, stop immediately:
- Do not write a proposal. Do not write a full verdict.
- Output exactly one line:
Skip — [core domain, e.g. "Odoo/ERP migration"], outside profile (Python/RAG/LLM focus). - Watch for jobs that list a familiar tool (e.g. "Python") as a secondary skill but the actual work is a platform/domain Caleb doesn't do (e.g. Odoo, SAP, Salesforce). The core task decides the fit, not one line in the skills list.
If the job genuinely mixes in-scope and out-of-scope work, don't skip — go to Step 3 and reflect that mix in the verdict.
Step 3 — Verdict
Compare the job's requirements against references/profile.md point by
point:
- Fit score (1-10), with the reasoning stated plainly, not just the number.
- Budget check: is it reasonable for the described scope? Flag if it's below what similar work should pay.
- Client quality signals: use the Client Information block if present — low review count + high spend, no payment verification, or a first-time client are things to flag, not dealbreakers on their own.
- Red flags: vague scope, no budget stated, unrealistic timeline, scope creep in the description.
- Recommendation: Apply / Apply with conditions / Skip.
Never credit Caleb with a skill or experience references/profile.md
doesn't support, even if it would make the proposal stronger. State a
partial fit plainly instead of framing it as a strong one.
Step 4 — Proposal (only if verdict is Apply or Apply with conditions)
Use assets/proposal_template.md as the skeleton.
- Read the full job listing before drafting — placeholders depend on real details from the post, not generic filler.
- Every
[placeholder]gets replaced with something real, from the job post orreferences/profile.md. None survive into the final text. - Delete the
## Sectionheaders from the template before presenting — the output is a flowing short proposal, not a labeled list. - Pick 1-2 portfolio projects that most resemble this specific job's deliverables, not the flashiest ones in general.
- Answer the job's specific asks directly (stack, timeline, deliverables) — no generic value statements.
- Closing questions target real ambiguities in the post (unclear scope, missing budget detail, unspecified tools) — not filler questions.
- If the listing includes Screening Questions, answer them directly and briefly, using only real information from the profile.
- Tone: short and direct. No emojis. No unnecessary exclamation marks. Target 80-120 words for the proposal body.
Step 5 — How to win it (only when there's a proposal)
Add 2-3 tactics specific to this job, e.g.:
- Apply fast if the post is recent
- Ask a sharp scoping question instead of quoting blind
- Reference the closest matching portfolio project directly
- If budget is a range, ask where in the range before committing
Output format
When a verdict + proposal is generated, use this structure, plain text, no emojis:
VERDICT — Fit X/10
[1-2 lines: why, budget assessment, red flags if any]
Recommendation: [Apply / Apply with conditions]
PROPOSAL
[ready-to-send text]
HOW TO WIN IT
[2-3 specific tactics]
When skipping (Step 2), output only the one-line skip — nothing else.
相关技能
把自然语言描述转为结构化 JSON,并由 mcp-diagram-generator MCP 服务生成 Draw.io、Mermaid 或 Excalidraw 图表文件。
以 AI 机器人身份加入视频会议,提供语音、虚拟形象与屏幕共享四种模式。
通过托管的 OAuth GraphQL 接口查询与管理 Linear 的 issue、项目、团队、周期、标签和评论。
通过一次 REST API 调用,向 10 个社交平台发布视频、图片、文字与文档。
诊断生产力系统反复失效的根因,给出最小干预——容量测算、瓶颈定位、可靠的本地记录。