How to write a memo with AI that people actually read

Most memos fail in the first sentence. They open with throat-clearing ("As we discussed in the last meeting, there are several factors to consider...") and make the reader dig for the point. A memo has one job: state a recommendation and give just enough reasoning to trust it. AI can get you a structured first draft of that in minutes, but only if you stop asking it to "write a memo" and start handing it a structure plus the facts.

This is the workflow I use: a paste-ready skeleton, the prompt scaffolding that keeps the AI honest, and the two edits you have to do by hand because the model will always get them wrong.

What a memo actually is

A decision memo is not a report and not an essay. It is a recommendation with its reasoning attached, written so a busy person can read the first line, get the answer, and decide whether to read the rest. That single constraint, answer first, is what separates a memo people act on from a memo people skim and forget.

Everything else follows from it. The structure exists to support a recommendation, not to build suspense toward one. If your reader has to reach paragraph four to learn what you want, you have written a mystery, not a memo.

The skeleton: six parts, recommendation first

Here is the structure I make every memo draft conform to. Paste it and fill it in:

# Memo: [one-line subject]
To: [reader] / From: [you] / Date: [date]

## Recommendation
The answer, in one sentence. What you want the reader to
decide or approve. No preamble.

## Context
Why this is on the table now, in 2-3 sentences. The
trigger, and the single fact that frames the decision.

## Options considered
The 2-3 real choices, including the one you are NOT
recommending. A memo with one option is a press release.

## Reasoning
Why the recommended option beats the others. Tie each
point to the context. This is the part that earns trust.

## Risks
What could go wrong with the recommendation, and your
mitigation for each. Concrete consequences, not "there
may be challenges."

## Decision needed
The specific ask: who needs to approve what, by when.

The order is the whole trick. Recommendation at the top means a reader who trusts you can stop after one line. The Options and Risks sections force honesty: if you cannot name an option you rejected or a risk you carry, you are selling, not advising.

The AI drafting loop

The mistake is asking an AI to "write a memo about X." It will produce generic corporate prose because it has none of your actual context. The fix is to feed it the raw facts and constrain it to the structure. Step one: gather the inputs the memo is about, the decision, the options, whatever numbers or notes you have, and paste them in. Step two: give it the skeleton and this prompt:

You are drafting a decision memo. I'm giving you the raw
context below. Fill in this exact skeleton:
Recommendation / Context / Options considered /
Reasoning / Risks / Decision needed.

Hard rules:
- Put the Recommendation in ONE sentence at the very top.
  No preamble, no "this memo will discuss."
- Use ONLY facts I provide. If a claim needs a number I
  didn't give you, write [NEEDS DATA] instead of inventing.
- Options considered must include at least one option you
  are NOT recommending, stated fairly.
- Each risk must be a concrete consequence plus a
  mitigation, not a vague "there may be challenges."
- No hedge words: "explore", "consider", "leverage",
  "potentially". Write commitments.

Output Markdown only.

--- RAW CONTEXT ---
[paste your notes, numbers, options here]

What comes back is a real first draft, structured, answer-first, and flagged where it lacks data. That is the 70% an AI is genuinely good at: organizing your raw notes into the shape of a decision.

The two edits the AI will always get wrong

The last 30% is the whole game, and it is two specific operations the model cannot do for you.

1. Un-bury the recommendation. Even when you tell it not to, an AI drifts back toward building up to a conclusion, because most of its training text is structured that way. Read the draft and ask: does the reader get the answer in the first sentence? If the recommendation has crept down into paragraph two, pull it back to the top and cut whatever throat-clearing preceded it. A memo is not a story with a reveal.

2. Restore the bite in the risks. Models smooth risks into soft, plausible-sounding language: "there may be some adoption challenges." That sentence cannot be acted on. Rewrite each risk as a concrete consequence and a mitigation: "If sales doesn't adopt the new flow by Q3, we lose the projected $40k uplift; mitigate by making the old flow read-only on August 1." Real risks have edges. The AI sands them off; your job is to put them back.

Why a memo is a document, not slides

A memo is reasoning, and reasoning lives in sentences. The moment you compress a decision into bullet fragments in 16:9 boxes, you lose the connective tissue, the because that links a claim to its support. A memo that says only "Option B: faster, cheaper, scalable" is three adjectives, not an argument. The reader cannot tell whether you reasoned your way there or just liked the vibe.

So write the memo as a document. If you later need to walk a room through the decision, render the same source as a deck for the presentation, but keep the canonical memo as readable prose. This is the same principle behind writing a strategy doc as prose first and only rendering the slide view later: the argument is the source, the format is a view of it.

Where Plain fits

This is the model Plain is built on for documents. You write the memo as Markdown, or describe the decision to the AI and edit the draft it returns, and Plain renders it as a clean web document you share as a link instead of an email attachment. A few things that follow from a memo being a web page rather than a .docx:

The link is always the current version, so when the decision evolves you update the source and everyone who opens the link sees it, no "memo_final_v3" in five inboxes. When the decision rests on data, you can embed the chart or table inline rather than describing it in prose. And because the source stays plain text, the memo diffs cleanly if you revise it, so the reasoning has a history. When a receiver genuinely needs the file, Plain exports to .docx as a downgrade, while the live link stays the version of record. The underlying idea is that the document is a link, not a file you attach and lose track of.

The short version

To write a memo with AI: hand it the six-part skeleton (Recommendation, Context, Options, Reasoning, Risks, Decision needed) and your raw context, constrain it to your facts and an answer-first structure, and let it produce the draft. Then do the two edits only you can do: pull the recommendation to the first line, and give every risk a concrete consequence and mitigation. Keep it to a page, write it as a document not slides, and deliver it as a link so it stays one current version.

The point is not that AI writes your memo. It cannot decide what you recommend, that is the part with your judgment in it. The point is that AI turns your raw notes into a structured, answer-first draft in minutes, so your time goes to the recommendation and the risks, the two things that actually determine whether anyone acts on it.