Inbox triage: reading before writing.

Why triage comes first, and why read-only CORE

You are about to do your first real agentic work on your inbox. The job is not to reply. It is to read — to reduce a label full of messages to a short, honest digest you can act on quickly. That is called triage, and doing it well is the highest-leverage thing an email agent does in a student’s life.

Triage comes first for two reasons that matter equally.

The first is skill. If you cannot direct an agent to categorize a day’s mail honestly — if you cannot write a prompt that produces “three messages that need a reply today, four FYIs, six newsletters, two I think are fine to archive” — you cannot direct it to draft replies honestly either. Categorize is the second move in the four-move flow. Every draft you write downstream sits on top of a categorization. A weak triage produces confident drafts on the wrong threads.

The second is safety. Triage is the one email-agent job that does not produce any outbound action. The agent reads; the agent summarizes; no message leaves your account. If you are going to misunderstand the agent’s judgment — and on the first day, you will — misunderstanding it in a read-only task is free. Misunderstanding it in a draft-and-send task costs you a message to a real person. You install the habit where the stakes are smallest, then carry it forward.

The rule for this lesson is simple and uniform: the agent has read-only access to the agent-access label. It writes nothing. It sends nothing. It does not create drafts, does not archive, does not mark as read, does not label. You are going to check this three times in the setup — when you configure the permissions, when you write the prompt, and when you read the digest. The first day with a new agent is the day to overdo this.

What a good digest looks like CORE

A good triage digest is a small number of clean categories, each with a few short lines, each line identifying a specific thread well enough that you can find it in under ten seconds.

Here is the shape of a triage digest that works. The categories and the order matter.

Reply needed today. One to five entries. For each: the sender’s relationship to you (not their name — the relationship, for your private notes), a one-sentence summary of the thread’s current state, and what a reply would roughly need to address. Example: “Thread with my teacher about the late assignment. She has asked whether I will submit Friday or Monday; reply needs one or the other.”

Reply needed this week, not today. One to five entries. Same shape. These are commitments you owe a response on, but the thread can wait a few days.

FYIs. Informational messages you should know about but do not need to reply to. One line each. Example: “Library hours shortened for the rest of April. No action needed.”

Archivable. Newsletters, confirmations, promotional emails that do not need to live in your triage workflow. The agent suggests archiving; you decide. Never: the agent archives for you.

Flagged as unclear. Anything the agent is not confident about. This category is required. If the agent has read ten messages and returns no flags, be suspicious — either the agent is pretending to certainty, or you have an unusually clean inbox. Ask it to name its two least-confident categorizations anyway.

Missed by the fence. If the agent notices a thread is referencing something it cannot see — a conversation that clearly started outside the agent-access label, a reply to an earlier message not in its scope — the agent names that missing context rather than inventing it. This is an honesty check.

What a bad triage digest looks like: a paragraph-long re-description of each email, written in a summarizing voice, with no categories. That is the agent being verbose instead of useful. Redirect it: “Give me five categories in the shape I just described, one line per entry, not paragraphs.” A good agent corrects fast.

The triage prompt pattern CORE

A triage prompt has five parts. Keep them in this order the first few times; once the pattern is internal, you can compress.

1. What you want done. Plain and specific. “Triage my agent-access label as it stands right now. Produce a daily digest.”

2. The categories you want. Name them. “Use these categories, in this order: Reply needed today, Reply needed this week, FYIs, Archivable, Flagged as unclear, Missed by the fence.” Do not assume the agent knows your categories. Every agent’s default categorizer is slightly different.

3. The shape of each entry. “For each entry, give me one line: sender relationship, one-sentence thread state, and for reply-needed entries, what the reply would address.” Shape matters. Without shape, the agent will give you paragraphs.

4. The honesty rails. “If you are not confident about a categorization, put the thread in Flagged as unclear and name your reason. If you see a thread that references material outside this label, list it under Missed by the fence. Do not invent subjects, senders, or thread content.”

5. The constraint. “Do not draft replies. Do not mark anything read. Do not archive. Read-only triage only.”

This is a lot of scaffolding for a short job. That is the point. The first day you run triage on your real mail, you want the agent to know exactly what you expect and exactly what it may not do. Over time, your prompt will shorten — most students end up with a two-sentence version saved as a snippet after a week. You do not start there. You start with scaffolding.

One thing not in the pattern: do not ask the agent to rank the replies-needed entries by “importance.” Importance is your call, not the agent’s; the categorization already tells you what needs your attention. Ranking is where the agent starts making judgment calls about your life and your relationships. You keep those.

Auditing the digest CORE

The digest arrives. Now audit it. The audit is the equivalent of the “review every diff” habit from Module 3, applied to a categorization.

Four audit checks, in order.

Count. Are there as many entries across all categories as there are messages in the label? Open the label, count. If the digest has fewer entries than there are messages, the agent dropped something. Ask it directly: “Are there any messages in the label you did not include in the digest? List them.” Good agents will admit to this. Tonight is one of the two or three times in the module where “ask the agent about its own completeness” is worth doing explicitly.

Placement. For each entry, ask: does the category match the message? Open the actual thread for two or three entries — especially the ones in Reply needed today — and confirm the agent’s one-line summary is what the thread actually says. Mis-placement is the most common triage error, and it is almost always in the direction of making the thread sound more urgent than it is.

Invention. Does any entry reference a sender, a subject, or a thread state that does not match reality? An agent with read access to real messages will hallucinate much less than one without — but it will still occasionally invent something plausible, especially when a real message is short or ambiguous. If you see an entry that rings wrong, open the thread.

Missing context. If the agent placed something under Missed by the fence, good — it is being honest. If the agent did not flag anything as missing context but you know from your own memory that one of the threads in the label references an earlier conversation not in scope, the agent missed the missingness. Add the item to your mental note.

After the audit, you will know two things. First, whether the digest is usable — whether you can act on it as-is or whether you need to re-prompt. Second, how reliable this agent’s triage is at this scope. The second answer is the one that matters long-term; it tells you whether to widen the access surface in the next module or keep it where it is.

Where to run triage: the Cowork tab is home base CORE

The Cowork tab in the Claude desktop app is the home base for Module 5 — and for triage specifically. It is the right shape for the work: a session you open, a conversation that runs the prompt, a digest you can read and follow up on (“okay, tell me more about the thread with my teacher”) before you move into drafting in 5.3. Chat-shaped. Document-oriented. The agent runs in a cloud VM and reaches your Gmail through the MCP connector you set up in Lesson 5.1. You spend most of your time looking at a conversation panel.

For students with terminal comfort, the Claude Code CLI is an Optional Advanced alternative — useful when you want triage to be scriptable and repeatable. A prompt you check into a small repo, re-run Monday through Friday, and maintain over time. The CLI path produces a digest file you can commit to git and re-run with one command tomorrow. It comes back into its own in Modules 6 and 8 where scheduled and orchestrated work runs from a folder. For this lesson, the Cowork tab is enough; the CLI is offered as a collapsed Optional Advanced sidebar in the recipe below if you want to try the scriptable path.

One honest note on picking a model: triage is a job where a smaller, cheaper model is often fine. You are asking for categorization, not creative writing. If you have a local-model option and the content in the label is sensitive — a parent’s small-business mail, for instance — this is a good lesson to try the local model and see whether it holds up. The capability weight for triage is smaller than for drafting.