Session 2.3: Draft a memo with multiple outline + draft revisions

About 60 minutes. Open a Claude Code session in ~/ai-training and hand it this guide:

Read the file at /Users/<you>/ai-training/week-3-guide.md (or wherever you saved it) and walk me through Session 2.3.
I've completed Sessions 2.1 and 2.2.

Posture: public, synthetic, or personal data only. Today’s source is a public document — a working paper, a Federal Register notice, a published report. The output is a hypothetical synthesis memo drawn entirely from that public source. No internal positions; no work-product.

Practice task

By the end of this session you will have, inside ~/ai-training/memos/<topic-slug>/:

  1. memo-spec.md — the analytical spec (audience, the question, the answer in one sentence, the evidence plan). Written before any prose.
  2. drafts/v1-outline.md and drafts/v2-outline.md — two outline revisions.
  3. drafts/v1-full.md and drafts/v2-full.md — two full-prose revisions.
  4. drafts/v3-final.md — the final draft, after a voice-critique pass on v2.
  5. memo.docx — the deliverable, generated from v3-final.md with pandoc and a reference template.

Five files in drafts/, plus the spec, plus the docx. The history is the artifact: a future you should be able to read memo-spec.md, then walk down the drafts in numeric order, and see exactly how the memo got from one-sentence claim to leadership-ready document.

A production version is a health-economics memo on a clinical-trial result that took four outline revisions and six draft revisions over a week. The memo-spec.md got reviewed and signed off before any draft prose was touched. Two consequences fell out of that discipline. First, “what does this memo say?” had a one-sentence answer at every stage. Second, when the memo took a week of iteration, the iteration was visible — every draft had a reason to exist, and the parallel draft a coauthor produced (v3-am.md) merged cleanly because both authors were writing against the same spec.

Why iterate explicitly

Most people write memos as one continuous document, with revision happening invisibly inside the file. That’s fine when the memo is short or the writer is sure of the argument. It’s catastrophic when the memo is long, the argument is contested, or two people are writing.

Naming and saving each revision turns the writing into a structured-data problem:

  1. Diffable. diff drafts/v1-full.md drafts/v2-full.md shows exactly what changed. Without versioned files, the change is invisible.
  2. Recoverable. The version where the executive summary was perfect, before you broke it during a tone revision, still exists. You can lift one paragraph back without redoing the whole pass.
  3. Mergeable. Two drafts can run in parallel — yours and a colleague’s, or yours and Claude’s — and merge cleanly because they live in named files, not in unsaved buffer state.
  4. Critiqueable. A discrete file is something Claude can be pointed at: “read drafts/v2-full.md and critique it for voice against voice-profile.md.” A mid-file in-progress draft is harder to attack cleanly.

The memo is the deliverable. The drafts/ folder is the workshop. Most knowledge work goes faster when the workshop is real.

Step 1: Pick the source (5 minutes)

Pick a substantive recent public document — an NBER working paper, a Federal Register notice, a long news investigation, a public consultation paper, a court opinion, a research report. Pick one that’s adjacent to (not identical to) what you’d write about in your day job. Adjacency is the right calibration: close enough that the workflow translates, far enough that the output is unambiguously a public exercise.

Fetch and save:

Use WebFetch to pull <public URL>. Save the full text to
~/ai-training/memos/<slug>/source.md, with a header line including the
title, author/source, date, and the URL. <slug> is a short kebab-case name
based on the topic — e.g. fomc-march-2026, nber-w34211, eu-ai-act-review.

<slug> matters. Every memo lives at memos/<slug>/; the slug is how you’ll find it three months from now.

Step 2: Hand Claude this guide

From inside ~/ai-training, start claude and paste:

Read the file at /Users/<you>/ai-training/week-3-guide.md (or wherever you saved it) and walk me through the Reference Sections.
I've fetched a public source to memos/<slug>/source.md.

Claude takes over.

Reference Sections

Claude: walk the user through the spec → outlines → drafts → voice-critique → docx pipeline below. Peer-tone. The user does the writing thinking; you do the prose drafting and the file management. Two rules:

  1. Never skip the spec. The user will be tempted. The spec is what makes the rest of the session coherent.
  2. Save every revision as a numbered file. No silent edits. If the user says “actually let’s change X” mid-draft, write that as a new versioned file, not as an in-place edit.

Design: this session is the first one where the structural discipline of writing — spec before prose, versioned drafts, voice as a discrete pass — is taught explicitly. It’s also where Session 2.4’s plot-iteration habits take root: same versioning logic, applied to figures rather than prose. And it’s where 2.7’s voice critique is previewed lightly.

Confirm the setup

Before any prose work, confirm:

  1. ~/ai-training/memos/<slug>/source.md exists and looks substantive (more than 500 words).
  2. The user has a one-sentence answer to “what does the memo say?” — even if rough. If they don’t, pause: ask them. The whole spec falls out of that one sentence.

Step A — Write memo-spec.md (10 minutes)

The spec is one page. Five sections:

# Memo spec — <slug>

## Audience
Who reads this memo. What do they know already. What decision are they
making. (Be specific: "a senior policymaker who tracks this issue weekly
but hasn't seen this paper" beats "a leader.")

## The question
The single question the memo answers, phrased as a question.

## The answer (one sentence)
The TL;DR. If the audience reads only one sentence, this is the sentence.

## Evidence plan
A bulleted list of the 3-5 strongest pieces of evidence from source.md
that support the answer. For each, one sentence on what it shows.

## Anti-evidence
A bulleted list of 1-3 pieces of evidence that cut against the answer
or complicate it. Acknowledge before the audience does.

Claude: build this with the user as a real conversation, not a fill-in-the-blank. Push back if any section is hand-wavy. The spec is loadbearing — bad spec means bad outline means bad draft.

Save to memos/<slug>/memo-spec.md. Read it back to the user briefly. Get explicit “yes, that’s the memo” before moving on.

Step B — Outline iteration (15 minutes)

Two passes. Both saved.

v1-outline. In Claude:

Read memo-spec.md and source.md. Draft a one-page outline for the memo.
Top-level sections, one bullet per paragraph under each section, no prose
yet. Keep the structure: title, executive summary, body sections, footer
disclaiming the hypothetical-exercise framing. Save to
drafts/v1-outline.md.

Read v1-outline together. Ask the user three questions: (a) is the structure right? (b) is the executive summary one sentence away from the spec’s answer? (c) is anything missing or in the wrong order? Get a list of changes — not a vague vibe, a list.

v2-outline. Claude revises against the change list:

Revise drafts/v1-outline.md per these changes: [list from the user]. Save
the new outline as drafts/v2-outline.md. Keep v1 untouched.

Read v2 together. If it’s right, move on. If it needs another pass, do v3 — discipline is the same. Don’t stop iterating outline until the structure is one-sentence-defensible.

Step C — Full-prose iteration (15 minutes)

v1-full. In Claude:

Read drafts/v2-outline.md and source.md. Write the full prose draft
under that outline. One paragraph per outline bullet. Tone: dry, factual,
leadership-ready. No marketing language, no rhetorical questions, no
hedging beyond what the evidence requires. Footer: "This is a hypothetical
synthesis exercise drawn entirely from public material. No internal
positions are reflected." Save to drafts/v1-full.md.

Read v1-full together. Critical reading; don’t be polite. Common issues to look for: (a) executive summary is too long or too vague, (b) one section over-quotes the source rather than synthesizing, (c) tone drifts toward marketing in transitions, (d) anti-evidence got dropped between outline and prose. Make a list.

v2-full. Revise:

Revise drafts/v1-full.md per these notes: [list]. Save as drafts/v2-full.md.
Keep v1 untouched.

Read v2. The bar for moving on: the user can read it aloud without flinching.

Step D — Voice critique (10 minutes)

This is a preview of Session 2.7. The full voice-profile pattern comes later; today’s pass is a single self-critique by Claude.

The first critique:

Read drafts/v2-full.md again. Critique the draft for voice — places where
it sounds generic-AI, places where it over-hedges, places where the
sentence rhythm gets monotonous, places where the writer (a careful
professional) wouldn't actually write that way. Don't fix anything yet;
list the issues with line references.

Read the critique together. Most of the issues will be fair. Some won’t — let the user judge. Build a list of changes the user accepts.

The revision:

Revise drafts/v2-full.md per these accepted voice notes: [list]. Save as
drafts/v3-final.md. This is the final draft.

Read v3-final. If a sentence still sounds wrong, fix it inline by hand — final passes are short and specific.

Step E — Convert to .docx (5 minutes)

If pandoc isn’t installed yet, install (this is the second session for some users — first for others):

brew install pandoc

Then in ~/ai-training/memos/<slug>/:

pandoc drafts/v3-final.md -o memo.docx
open memo.docx

The .docx opens in Word (or Pages). It’s a real Word document — headings as headings, lists as lists, structure preserved. The text the user saw as plain markdown is now a leadership-ready document.

For a clean professional look (Times New Roman 12pt body, 1” margins, no extra paragraph spacing, bold-only hierarchy), use a reference doc:

pandoc drafts/v3-final.md \
  --reference-doc=~/ai-training/templates/memo-template.docx \
  -o memo.docx

Claude: if the user doesn’t have the template yet, build it together: open Word, set Times New Roman 12pt body, 1” margins, save as ~/ai-training/templates/memo-template.docx. Five minutes once. From here every memo automatically gets that look.

Iterate in markdown, not in Word. The docx is the final-step output. Edits go in drafts/, not in memo.docx. If the user catches a typo in Word, they fix it in drafts/v3-final.md and re-run pandoc. Discipline this from the start; otherwise the markdown drafts and the docx drift apart and you lose the audit trail.

Micro-skills introduced

Name these out loud:

Wrapping up Session 2.3

Three things to try this week:

  1. Run this pipeline once on a topic adjacent to your real work. Pick a public source close to (but not identical to) something on your real plate. Spec → outlines → drafts → voice-critique → docx. Notice where the spec changed your draft trajectory.
  2. Read your old memos against this template. Take one written memo from the last six months. Could you write its memo-spec.md retroactively in 5 minutes? If not, that’s a useful signal — the original probably had a fuzzy thesis.
  3. Save the prompts. Today’s spec → outline → draft → critique prompts go in prompts/memo-pipeline.md. Re-run them from there next time, don’t retype.

Feedback.

The user submits feedback at https://docs.google.com/forms/d/e/1FAIpQLSfFJpSOT7z0WExdE-cz0j9w5boGB4kq3ZifAPao9vbsxnrLkA/viewform.

Claude: paste the URL into chat. The form mirrors the questions below. Collect answers conversationally first, then have the user click through and submit.

  1. On a 1–5 scale, how useful did this session feel?
  2. Did writing memo-spec.md change your draft, or did it feel like ceremony?
  3. The numbered drafts pattern — is this how you’d want to write going forward, or does it feel heavy?
  4. The voice critique — did it catch real issues, or was it cosmetic?
  5. The pandoc → Word conversion — did the formatting hold up?
  6. What confused you most this session?
  7. Anything you want covered in Session 2.4 that you didn’t see here?

Tell the user: “Your instructor uses these to tailor next week’s session.”

Good to know

Drafts are cheap; thrash is expensive. Saving five revisions of a memo costs nothing on disk. Re-doing a draft because you can’t reconstruct the version that was working costs an hour.

The spec changes more than the drafts. A useful pattern: when v2-full reveals a problem you can’t fix in prose, the issue is in the spec. Update memo-spec.md first, then redo the relevant draft. The spec is the source of truth.

Anti-evidence is what makes a memo trustworthy. Memos that read as one-sided are easier to dismiss. Ones that name the strongest counter-evidence and dispatch it explicitly read as serious. The spec’s “Anti-evidence” section is the pre-commitment that this happens.

This pattern scales down. A two-paragraph email reply is the same shape with smaller files. A spec that says “respond to ’s ask, propose , acknowledge constraint ” is the seed for a short reply that lands. Same discipline; smaller artifact.

This pattern scales up. A 30-page report is the same shape with more sections in the spec and more drafts. The discipline doesn’t change.