Track 02 · Workflows · Lesson 01

SAMPLE LESSON

Inbox Triage Workflow: From Chaos to Action List in 5 Minutes

What you'll be able to do after this

→Run a repeatable morning triage that turns your inbox into a short prioritized action list
→Classify any batch of emails into act, defer, read, and archive — in under 60 seconds
→Generate send-ready draft replies for your highest-priority emails with one prompt
→Extract all action items with owners and due dates from any email thread
→Route each step to the right model (Claude vs Gemini) to get the best output

The Problem: A Messy Inbox Has No Priority Order

Every morning, most knowledge workers open their inbox and face the same problem: 40, 80, 120 messages — all presented in reverse chronological order, with no signal about what actually matters. The inbox is organized by arrival time, not by importance to you.

The result is a cognitive tax you pay before 9 AM. You re-read the same messages trying to decide whether to act on them. You start drafts and abandon them. You defer things without a plan. You forget to follow up. Each unprocessed email is a micro-decision you haven't made — and it costs attention even when you're not looking at it.

This workflow solves that. Every morning, in five minutes, you turn your inbox into a short prioritized action list: a handful of things to act on today, a clear defer queue, and everything else archived or filed. Not by spending more time on email — by processing it deliberately, with AI, once.

The Four-Step Workflow

This workflow runs in sequence every morning. Each step has a dedicated prompt — you paste input, get output, and move to the next step. Total time: 5 minutes.

Paste

Collect your inbox

Copy the last 24–48h of emails: sender, subject, and the first line of each. Paste them into a single text block. This is your raw material — takes about 60 seconds.

Classify

Triage with Prompt 1

Run the Triage Scan prompt. Every email gets categorized into ACT TODAY, DEFER, READ, or ARCHIVE. Bulk-archive the ARCHIVE list immediately. Your signal-to-noise ratio just dropped by 60–80%.

Draft

Reply with Prompt 2

Take the ACT TODAY emails, paste them in full, and run the Draft Reply prompt. You get a send-ready draft for each. Review the 2–3 highest-stakes ones; send the rest as-is or with minor edits.

Extract

Pull tasks with Prompt 3

Paste any emails that contain commitments, requests, or decisions into the Action Items prompt. Out comes a table: task, owner, due date, priority. Copy it into your task manager or doc. Inbox processing is complete.

The Three Prompts

Save each prompt to your prompt library (see the previous lesson) and fill in the context sections once. After that, each morning you only paste new email content — the context is already there.

Prompt 1 — Triage & Classify

Route to: Claude Sonnet 5 · Input: sender + subject + first line for each email

COPY — inbox-triage-classify.md

You are a chief of staff helping me process my inbox. I will give you a list of emails (sender + subject + first line).

Categorize each email into exactly one category:
- ACT TODAY: requires a response or decision within 24h
- DEFER: requires a response, but not today — estimate a by-when date
- READ: informational, no action needed — worth reading but no reply required
- ARCHIVE: newsletters, automated emails, FYIs, completed threads — no action, no reading needed

Output format:
ACT TODAY:
- [Sender] | [Subject] | [Why it needs action] | [Urgency: HIGH/MEDIUM]

DEFER:
- [Sender] | [Subject] | [By when] | [What's needed]

READ:
- [Sender] | [Subject] | [Why worth reading, 1 sentence]

ARCHIVE:
- [Sender] | [Subject]

Context about me: [add 2–3 sentences about your role, company, and who your most important contacts are]

Email list:
[PASTE SENDER + SUBJECT + FIRST LINE FOR EACH EMAIL]

Prompt 2 — Draft Reply

Route to: Claude Sonnet 5 · Input: full text of ACT TODAY emails

COPY — inbox-draft-reply.md

You are drafting email responses for me. For each email below, write a reply I can send with minor edits or no edits.

My voice: [describe in 3–5 words — e.g. "direct, brief, action-focused" or "warm but efficient"]
My role: [your title and company in one sentence]
Format: subject line first, then the draft. Max 4 sentences per reply unless the email requires more.

Rules:
- Do not start with "I hope this finds you well" or any version of it
- End every reply with a clear next step or question
- If the email is from a VIP (I'll mark them), be slightly warmer in tone
- If you need information I didn't provide, note it with [MISSING: what you need] rather than making it up

Emails to reply to:
[PASTE EACH EMAIL IN FULL — SEPARATE WITH "---"]

Prompt 3 — Extract Action Items

Route to: Claude Sonnet 5 · Input: any emails containing commitments, requests, or decisions

COPY — inbox-extract-actions.md

You are extracting action items from a set of email threads.

For each email, identify every explicit or implied task, commitment, or decision, then output a table:

| Action | Owner | Due Date | Source (sender) | Priority |
|--------|-------|----------|-----------------|----------|

Rules:
- Owner is whoever must take the action — could be me, the sender, or a third party mentioned in the thread
- If no explicit due date, infer from urgency language: "ASAP" = today, "by end of week" or "soon" = this Friday, no signal = next week
- Priority: HIGH (blocks others or has a hard external deadline), MEDIUM (important but flexible), LOW (nice-to-have or informational)
- If an email contains no action items, skip it
- Include one row per action — if one email has three tasks, write three rows

Today's date: [DATE]
My name / role: [so the model can correctly assign owner = "me" vs owner = someone else]

Email threads:
[PASTE FULL EMAIL THREADS HERE — SEPARATE WITH "---"]

Routing Note: Claude vs Gemini for Each Step

Not every step benefits from the same model. The inbox triage workflow runs almost entirely on Claude — but there is one scenario where switching to Gemini pays off significantly.

Step 1 — ClassifyCLAUDE

Default: Claude Sonnet 5

Triage is a judgment task, not a reading task. Claude understands relationship context — who matters to you, what urgency sounds like between the lines, when a CC should be treated as a direct ask. Gemini will process the same emails correctly most of the time, but Claude's nuanced understanding of social signals and priorities produces better categorization for knowledge work. Exception: if you're processing 80+ emails with long threads and hitting context limits, route the first triage pass to Gemini 3.6 Flash for its extended context window.

Step 2 — Draft ReplyCLAUDE — ALWAYS

Default: Claude Sonnet 5

Email drafting is exactly what Claude is built for. It produces prose that sounds genuinely human, maintains your specified voice consistently across multiple drafts in the same session, and catches when your brief is underspecified (e.g. if you asked it to commit to a date but didn't give one, it will flag it rather than invent it). Gemini can draft emails but the voice consistency is noticeably weaker — it tends to drift toward a generic professional tone after 2–3 emails in the same session.

Step 3 — Extract Action ItemsCLAUDE (default) / GEMINI (large threads)

Default: Either — Claude preferred

Both models handle structured extraction well. Claude is preferred when you have 5–10 emails with complex inter-thread commitments — it tracks cross-referenced ownership and infers implicit deadlines more reliably. Use Gemini when you have a single very long email thread (20+ messages) or when multiple large documents are attached — Gemini's extended context window means it won't lose information from early in the thread.

The general rule for this workflow: Claude is your output craftsman — use it for anything that requires judgment, voice, or structured precision at moderate scale. Switch to Gemini only when you're hitting context limits or processing large multimodal attachments (PDFs, scanned docs attached to emails). For most inboxes on most mornings, Claude Sonnet 5 handles all three prompts.

Make It a Daily Habit

The workflow is only valuable if it runs every day. Here is how to make it stick — and what to do in the first week to tune it to your inbox.

The 5-minute morning schedule

9:00 AM

Paste email batch

Select the last 24–48h of emails, copy sender + subject + first line for each. Paste into a text block. Block archived or already-handled emails if you track them separately.

9:01 AM

Run Prompt 1 — Classify

Paste into Claude Sonnet 5. 15 seconds. Bulk-archive everything in the ARCHIVE list. You now have a prioritized action list instead of a wall of messages.

9:02 AM

Run Prompt 2 — Draft replies

Paste the full text of your ACT TODAY emails. Review drafts for 60–90 seconds. Send the straightforward ones as-is; edit only the high-stakes ones.

9:04 AM

Run Prompt 3 — Extract tasks

Paste any emails with commitments or requests. Copy the action items table into your task manager or a daily notes doc. You now have owners and due dates, not just a vague sense of 'I need to follow up on this.'

9:05 AM

Close inbox

Everything has a status. Nothing is ambiguous. You're done with email for the morning. No ambient checking — open inbox again at 1 PM if your role requires it.

Tuning — what to fix in the first week

⚑ Miscategorized emails

Improve the 'Context about me' section in Prompt 1 — add your most important contacts by name and describe your role more specifically. More context → better categorization.

⚑ Drafts don't sound like me

Add a style example to Prompt 2. Paste one or two real sent emails you're happy with. Claude will match the pattern immediately.

⚑ Too many items in ACT TODAY

Add a filter: "Only put an email in ACT TODAY if I am the primary recipient and a response is genuinely required — not CC'd, not a courtesy reply." If ACT TODAY is still overflowing, that's the real problem to solve.

⚑ Takes longer than 5 minutes

You're editing too carefully. Trust the drafts. Spend review time only on the 2–3 highest-stakes emails. Send the rest as-is.

The compounding effect: After one week of tuning, the prompts are calibrated to your specific inbox and voice. After 30 days, the 5-minute routine is faster than opening your inbox manually used to be — and you've reclaimed the cognitive overhead that was previously spent deciding what to respond to. That's what an AI OS does: it converts recurring friction into a solved system that runs on autopilot.

What's Next

Build the rest of your AI OS

Upcoming lessons cover research pipelines, writing workflows, weekly reviews, agent design, and full automation. New content monthly.

Subscribe — $19/mo→

Cancel anytime · No contracts