Drafts, not sends.

Why the agent drafts, why you send CORE

In Lesson 5.2 the agent read and categorized. Nothing left your account. In this lesson the agent starts writing — in your voice, about your relationships, to people who know you. The stakes change. So does the rule that protects you.

Drafts, not sends. The agent writes the reply. The draft lands in your drafts folder. You open it, read it, adjust it, and click Send yourself. The agent does not click Send. The agent does not “send automatically for messages marked low-stakes.” The agent does not “send drafts it is confident about.” In this module there is no configuration, no setting, no prompt that lets the agent hit Send. If you find yourself looking for a shortcut around this rule, the shortcut is the mistake.

Why the bright line. Three reasons, in order.

First, the recipient does not know there was an agent. When your teacher opens a message from your account, she reads it as you. A wrong sentence in that message is a wrong sentence from you. There is no footnote that says “my agent drafted this” — and there should not be, because the relationship is yours, not the agent’s. The only way to preserve that relationship honestly is for a human who actually holds it to be the one who releases the message.

Second, fluency is not correctness. Language models are trained to produce plausible, fluent prose. An agent that can draft a reply to your cousin in a voice that sounds like yours can also draft a reply with an invented commitment (“I’ll bring dinner on Friday” — you never said that), an off-voice formality (“Dear Ms. Chen” — to someone you call by first name), or a smooth connector sentence that papers over a factual uncertainty (“I’ll get that to you by end of week” — you do not know if you can). Fluent drafts slip past a fast reader. That is the entire category of error this module is installing the habit against.

Third, the habit survives the tool. The specific agent you are using in April 2026 will be different by the end of the year. A more capable model next month will write better drafts. The norm does not move with the model’s capability. The human’s role is not conditional on how good the agent is; it is conditional on what the outbound message represents. Even if the drafts were always perfect, the human’s review would still be the point where a person’s words are affirmed as their own.

You will be tempted, a week into using this, to relax the rule “just for newsletters” or “just for routine confirmations.” Do not, in this module. After the module — with a month of lived experience, with your own data on what your specific agent gets right and where it reaches — you may choose which narrow cases deserve a relaxation. You do not earn that choice by shortcutting the habit now.

Giving the agent enough context to draft well CORE

A drafting prompt does more work than a triage prompt, because the agent has to know not just what to say but how you say it to this person. The three-context pattern gets you most of the way there.

1. Thread context. What has been said in the thread so far. The agent can read this itself from the thread — that is what inbox access is for — but you often want to direct it: “Read the thread from my teacher about the late assignment. The current state is she has asked whether I will submit Friday or Monday.” You name what the thread is about so the agent does not wander.

2. Relationship context. The agent does not know the recipient the way you do. Tell it. “This is a teacher I have a good relationship with; I usually sign off with my first name; she has extended deadlines for me before; the tone is warm and professional but not formal.” Two sentences is enough. This is the single highest-leverage thing you can add to a drafting prompt, and the thing students forget most often.

3. Content direction. What the reply should actually do. “Tell her I’ll submit Friday. Thank her for the extension. Don’t over-apologize — I’ve already apologized in an earlier message; doing it again would ring false. Keep it to three or four sentences.” Tell the agent the shape, the substance, and what not to include. The “do not over-apologize” kind of instruction is something only you know.

After those three blocks, add two standing rails.

Honesty rail. “If you are not sure about a fact or commitment, put a bracketed placeholder — [DATE?] or [CONFIRM WITH MOM] — rather than inventing. Do not fill in plausible details.” The placeholder pattern is how you get the agent to flag its gaps instead of papering over them.

Draft-only rail. “Put this in my drafts folder. Do not send. Do not mark the parent thread as replied or read.” Every time. The rail is redundant with the OAuth scope, and it is still worth saying. Two-layer defense.

That is the pattern. For a tough thread you may write a longer prompt. For a short reply to a friend you may write two sentences. The pattern scales down; do not let it scale down below relationship context, which is the one block the agent cannot infer.

The review-before-send checklist CORE

You have a draft in your drafts folder. Before you click Send, run the five-check review. For the first week, do it out loud. After that, do it silently, but do it every time.

1. Voice. Does the draft sound like you to this person? Read the first and last sentence aloud. Would you open a message to this recipient this way? Would you close it this way? A common off-voice move is the agent opening with “I hope this email finds you well” to a friend. That is a professional email opening; it does not belong on a message to your cousin. Delete, replace with what you would actually say.

2. Facts. Is every factual claim in the draft something you know to be true? Dates, times, commitments, names of documents, amounts of money, plans with other people. Open the thread and check. Anything you cannot verify inside five seconds — correct it, delete it, or hold the draft and ask a human who would know.

3. Relationship fit. Is the level of formality, warmth, and directness right for this relationship? Does the draft over-apologize, under-apologize, or hit the right note? Does it over-explain (a common drafting tic when the agent is unsure of tone) or under-explain (when the agent assumed shared context the recipient does not actually have)? Adjust.

4. What is missing. Is there something you would normally include that the agent did not think to add? A specific callback to something the recipient said in an earlier message. A question you meant to ask. An acknowledgment of something going on in their life. Missing things are invisible — the draft reads fine without them, and yet the sent message would have been better with them. This is the check that most rewards going slow.

5. What is extra. Is there anything in the draft you would not say? Unnecessary filler, a platitude, a commitment you did not mean to make, an opening like “I wanted to reach out to let you know” that you never use. Cut. A shorter draft that is actually yours beats a longer one that sounds like an agent.

Five checks. A short pass per draft the first week; faster as the habit lands. If you cannot honestly say you did each check, you did not review the draft — you skimmed it. Skimming is the bypass.

Three drafting failure modes you will see this week CORE

Three failure modes account for most of the wrong sends students have after they relax the review discipline. Name them now so you recognize them on sight.

The smooth wrong claim. A sentence that is factually wrong but reads as if the writer knew what they were talking about. “I’ll be there by 4.” (You cannot be.) “The document is attached.” (It isn’t — the agent cannot attach things, and said so anyway.) “We already agreed to meet Tuesday.” (You did not.) The smoothness is the problem. A clunky wrong sentence makes you stop and reread; a smooth one carries you past the check.

Defense: the Facts check, applied to every factual claim without exception.

The off-voice reply. A draft that is competent prose but is not the kind of thing you would write. Too formal for a friend. Too casual for a work contact. Lightly alien in rhythm. The sentence structure is right but the voice is not yours. Recipients who know you will notice — not because they are analyzing, but because the message feels a step off. One off-voice reply is a small thing. Three of them in a row, to someone who knows you well, is the beginning of a relational confusion the agent will not repair.

Defense: the Voice check — read the first and last sentence aloud, and the Relationship fit check.

The over-committing sentence. A draft that, in service of sounding helpful, commits you to something you did not intend to commit to. “Happy to help with that anytime.” (“Anytime” is a commitment; did you mean it?) “Let me know if you need anything else and I’ll take care of it.” (“Anything else” and “take care of it” are both over-broad.) The helpful-sounding commitment is the agent’s instinct toward closing the email warmly, and it is the instinct that most often creates a problem in the thread that follows.

Defense: the What is extra check — hunt for open-ended commitments and constrain them to what you actually meant.

A student who has read this lesson once and recognizes these three patterns on sight has most of what the module is trying to install. The checklist catches them; the pattern-naming makes the checklist fast.

What to do when a draft needs three rounds CORE

Sometimes the first draft is mostly right and you can finish it quickly. Sometimes it is not. The question is when to re-prompt the agent and when to take the keyboard back yourself.

Re-prompt when the problem is a constraint you did not name. The draft is too long — tell the agent “keep it to four sentences.” The draft is too formal — tell it “this is a friend, first-name basis, warmer.” The draft missed context — paste in the earlier thread it missed, and ask again. A constraint you can name is a constraint the agent can follow; re-prompting is often faster than editing.

Re-prompt when the problem is a whole-structure issue. The draft buried the main point; the draft put the ask before the context the ask depends on; the draft started with a backstory that should be a sentence. Ask for a restructure — “lead with the ask, then one sentence of context.”

Take the keyboard back when the problem is voice or relationship, and the agent has already had one re-prompt to fix it. Voice is subtle. The agent may not be able to hit your specific voice even with good instruction, especially on a hard-to-write thread. If the second draft is still off, rewrite it yourself. Do not do three rounds of prompting on a draft you could have written by hand in less time. The agent is a tool, not a requirement.

The related question: when is it okay to use the agent’s draft as a scaffold and write over half of it by hand? Almost always. A student who rewrites a substantial portion of the agent’s draft is not failing the tool. They are using it for what it is good at — getting past the blank-page moment, sketching the structure, drafting the safe parts — and keeping the voice themselves. Treat the agent as a collaborator whose first pass you always revise. That is often the right level.