菜单
Use when drafting email replies for Michael, generating outbound emails, or applying his email communication patterns and style.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | email-reply-style |
| description | Use when drafting email replies for Michael, generating outbound emails, or applying his email communication patterns and style. |
When generating email replies or drafts for Michael, follow this style guide.
FIRST, determine if this is OUTBOUND or INBOUND:
Michael is sending TO someone, not responding:
is_reply: false in training dataFor OUTBOUND emails: Generate content Michael would SEND, not receive.
Michael is responding to someone else:
is_reply: true in training dataoriginal_message contains what they sentFor INBOUND emails: Respond to the thread context.
Michael uses Superhuman for automated actions. Recognize these:
When email is TO a weird hash address (like LBPWIYZTGQ...@unsub-ab.mktomail.com) with subject "Unsubscribe":
This is an unsubscribe request sent from Superhuman on behalf of michael@geteverything.aiALWAYS use the original_message to understand:
Example of proper context usage:
Email is MORE formal than messaging:
| Context | Length | Example |
|---|---|---|
| Simple acknowledgment | 1 line | "Thanks" or "Got it, thanks!" |
| Quick reply | 2-3 sentences | Answer + sign-off |
| Substantive reply | Paragraph(s) | Full response with context |
| Intro/meeting request | Short + links | Greeting, 1-2 sentences, links |
Quick Acknowledgment:
Thanks
Simple Reply:
Hey [Name],
[1-2 sentence response]
Best,
Michael
Reply with Links:
Hey [Name],
[Brief context]
Here's the event: https://luma.com/...
If you have questions: https://cal.com/everythingai/15min
Thanks
Substantive Reply:
Hey [Name],
[Main response - can be multiple paragraphs for complex topics]
[Optional: next steps or links]
Best,
Michael
When someone asks about timing:
When following up on a meeting:
When someone reports an error:
When asking about payment:
When confirming meeting times:
Before generating an email reply, retrieve similar emails from Michael's history:
llm similar michael-emails -c "Subject: Re: [subject]\n[original_message snippet]" -n 3
This returns JSON with similar emails and Michael's actual replies. Use these as few-shot examples to:
The embeddings are stored in ~/.llm.db and indexed via launchd (com.michael.email-index).
Full training examples with Michael's actual replies are in references/training-examples.md.
Each example includes: to_name, to_email, subject, timestamp, is_reply, original_message, and michael_reply.
(Add patterns for specific contacts as learned)
(Log corrections from Michael here to improve future replies)
Michael's operating + emotional coach. Operational mode — daily startup/shutdown, weekly review, pomodoro, inbox capture, daily notes (auto-loads in ~/ws/notes). Emotional/decision mode (Joe Hudson style) — use on "coach me"/"joe coach", stuck/looping/overthinking, harsh self- or other-judgment, a binary either/or decision that won't resolve, or fear, shame, loneliness, anxiety, burnout, or grief, when Michael wants to be met in a feeling rather than handed advice. Not for clinical crises (refer out).
Use when creating or processing Todoist tasks, triaging inbox items, doing daily task review, calibrating Todoist triage behavior, or turning corrections into reusable preferences. Routes to operations (CLI actions) vs calibrated triage (policy, context recovery, preference memory, evals). Trigger this whenever the user asks what to do with Todoist items, wants better task triage, or is refining how Todoist decisions should work.
Run real Deep Research across ChatGPT, Claude, and Gemini in parallel via the user's own logged-in browser (Chrome extension, zero API cost), save each original report to Notion, then synthesize. Use whenever the user wants a "deep dive", "deep research", a thorough multi-source investigation, or to research a topic across the models and compare what each finds. Drives the paid subscription products, NOT the API. NOT for single-fact lookups or ordinary web search — use web-search for those.
Use when setting up, auditing, or improving AI agent infrastructure in a repo — AGENTS.md/CLAUDE.md files, linters, architectural constraints, feedback loops, context tiering, agent specialization, or entropy management. Also triggers on "harness engineering", "agent-friendly repo", "make my repo work well with coding agents", "set up my repo for agents", or "why is my agent struggling".
Interactive Todoist triage with preference learning. Use when the user says "triage", "process my inbox", "clean up tasks", "triage my todoist", "file these captures", or mentions inbox zero. Also use when the user has a batch of raw items (voice notes, links, ideas) that need classifying and routing to Todoist projects or Obsidian. Runs an interactive confirm/correct loop that learns your routing preferences over time.
Use when the user asks to save session context, identify the current session or thread, create a resumable handoff, or prepare a Todoist/note summary that must include the working directory and session id. Works across Codex and Claude Code by detecting runtime-specific session identifiers and normalizing them into one summary.