/six-pager -- Amazon-style Narrative Memo and PRFAQ

Generate decision documents in the format Amazon has used since June 9, 2004 — when Bezos banned PowerPoint and required every meeting to open with a written narrative.

This skill produces the document. It does not replace the discipline. The original Amazon practice takes a week (draft → review → set aside → edit → final). The skill compresses the structure but preserves the prose-quality bar.

Why This Works

Three independent traditions converge on a single principle: writing is the thinking instrument, not the documentation of it.

• Bezos (Amazon, 2004): "There is no way to write a six-page narratively structured memo and not have clear thinking."
• Strunk (1918, restated by E.B. White 1959): "Vigorous writing is concise. A sentence should contain no unnecessary words, a paragraph no unnecessary sentences, for the same reason that a drawing should have no unnecessary lines and a machine no unnecessary parts." (Rule 17)
• Anthropic (Claude Code Best Practices): "For every line, ask: If I removed this, would Claude make a mistake? If not, remove it."

Same rule, three scales: line, sentence, instruction. This skill enforces all three.

When to Use This vs Sibling Skills

Natural pairing: Generate the document with /six-pager, then stress-test with /adversarial-review, then deliberate the recommendation with /council-review. That's the full Amazon-style decision pipeline.

Arguments

• Argument 1 (required): The topic, draft, or file path. One of:
  • A short topic statement: "Should we migrate the API to cursor-based pagination?"
  • A path to an existing draft: docs/pagination-decision-draft.md
  • A pasted block of context
• Mode flag (optional): memo (default for decisions) or prfaq (for launches). If omitted, the skill auto-classifies and asks to confirm.
• --silent-read: Adds a 20-30 minute "silent read" simulation to the output — produces an annotated version with margin questions a senior reviewer would write.
• --strunk-only: Skips structural generation; treats the input as an existing draft and runs only the Strunk + Anthropic prose audit. Useful for cleaning up an in-flight document.

Procedure

Phase 1: SCOPE

1. Parse flags from $ARGUMENTS.
2. Classify the input:
   • Topic — short statement, no draft. Generate from scratch.
   • File path — existing draft. Read it; treat as starting point.
   • Pasted text — treat as starting context.
3. If mode flag is missing:
   • Auto-classify: input contains "launch" / "ship" / "announce" / "new product" / "release" → suggest prfaq. Otherwise → suggest memo.
   • Confirm with the user before proceeding.
4. Confirm scope back to the user in one sentence: "Producing [memo|prfaq] for [topic]. [Optional: --silent-read enabled.]"

Phase 2: DRAFT

Mode: memo (6-pager)

Generate the six canonical sections. Each section has a specific purpose and a strict prose discipline:

1. Introduction (~0.5 page) — What is this document? Who is the audience? What is the ask? Frame the decision in one paragraph.
2. Goals (~0.5 page) — Specific, measurable outcomes. Each goal must be verifiable. Numbers required where possible.
3. Tenets (~0.5 page) — The principles this proposal optimizes for. No qualifiers (per Strunk's "avoid leeches" reminder). Tenets are unhedged commitments. Format: 1. [Tenet name]: [single declarative sentence].
4. State of the Business (~1.5 pages) — Current data. Honest. Concrete numbers, named systems, dated facts. Apply Strunk Rule 16 (definite, specific, concrete) ruthlessly.
5. Lessons Learned (~1 page) — What did past attempts teach? What is different this time? Each lesson cites a specific past event or experiment, not abstract principles.
6. Strategic Priorities (~2 pages) — The recommended action and rationale. Specific milestones, owners, timelines.
7. Appendix (unlimited) — Visuals, supporting data, alternative analyses considered and rejected (with reasons).

(V2) Premortem (required in Strategic Priorities or the Appendix). Assume it is 12 months later and this decision failed. Write the 3 most likely causes, working backward from the failure. This is forward-looking risk-finding that "Lessons Learned" (backward-looking) structurally misses — Kahneman's single highest-value decision technique, and it pairs with the verification: each premortem cause should map to a thing you'll monitor. Not generic "risks" — concrete failure scenarios.

Hard constraint: sections 1-6 combined must fit in 6 pages of 11pt body text. The skill renders a page count and warns if over.

Mode: prfaq (Press Release + FAQs)

Generate three sub-documents. Work backwards — write the press release first, as if the product already shipped.

1. Press Release (~1 page) — Customer-facing language. Benefit-led. Dated for the (hypothetical) launch day. Format follows Amazon's internal PR conventions:
   • Headline (one line, customer-outcome-oriented)
   • Sub-headline (one line, who-it's-for + key-benefit)
   • Lede paragraph (city, date, "Today X announced Y...")
   • Problem paragraph (what was hard before this product existed)
   • Solution paragraph (how the product solves the problem)
   • Customer quote (fictional but realistic)
   • Internal quote (from the team, on why this matters)
   • Call to action / availability
2. External FAQ (~1 page) — Questions a journalist or customer would ask:
   • What is this?
   • How is it different from [obvious alternative]?
   • How much does it cost?
   • When can I get it?
   • What does it not do?
3. Internal FAQ (~1-2 pages) — Questions Amazon-style leadership would ask:
   • Why now?
   • What does success look like? (specific metrics, specific dates)
   • What are the dependencies?
   • What is the failure mode?
   • What are we choosing not to build?
   • What would cause us to kill this?

The PRFAQ test: if you can't write a credible press release, the product doesn't exist yet — not in the form that matters.

Phase 3: CONSTRAIN

After Phase 2 produces a draft, measure it:

# rough page count: assume 450 words/page at 11pt
WORDS=$(wc -w < draft.md)
PAGES=$(echo "scale=1; $WORDS / 450" | bc)

If over budget:

• For memo: identify the longest section, propose 30pct cuts that preserve load-bearing content
• For prfaq: identify whichever FAQ sub-section ran long; propose merging or trimming questions

Never silently truncate. Always show the user what's being proposed for cuts and why.

Phase 4: AUDIT (Strunk + Anthropic)

Run the prose-quality audit on the draft. Each finding cites the rule:

Produce a structured audit report:

## Prose Audit (Strunk + Anthropic)

**Findings:** N critical, M important, K nit

### Critical (block the document)
- Line 23: passive voice in Strategic Priorities ("It was decided that...") → "We decided to..."
- Line 47: vague Goal ("improve performance") → demand a number
- Tenet 3 hedges ("We strive to be customer-obsessed") → cut "strive to" or cut the tenet

### Important (fix in next pass)
- 4 sentences in a row of 18-22 words in State of Business → vary
- Parallel-form break in Goals: 3 noun-phrase goals + 1 verb-phrase goal
- State of Business ¶2 buries its claim in the final sentence → lead the paragraph with it
- Overstatement in Introduction ("a massive, game-changing opportunity") → name the number or cut

### Nit
- "in order to" appears 3 times → "to"
- "various improvements" → name them or cut

Phase 5: PRESENT

Show the user:

1. Final document (or draft with audit annotations if --silent-read)
2. Page count vs budget
3. Audit summary (count of critical/important/nit findings)
4. Side-by-side: original draft vs proposed rewrites for any audit findings

Ask: "Apply the audit fixes? (yes / partial / no)"

If yes: apply fixes in place (surgical Edits if input was a file; show new full document otherwise). If partial: walk through each fix individually. If no: deliver the unaudited draft with a note that audit was skipped.

Phase 6: SAVE

Default save path: alongside the input file (if file input) or <cwd>/six-pager-<topic-slug>-<YYYY-MM-DD>.md.

If --silent-read was used, also produce <same-name>-annotated.md with margin questions in HTML comments or callout blocks.

The Silent-Read Simulation (--silent-read)

The original Amazon meeting protocol: distribute the document at the start of the meeting (NOT pre-read), 20-30 minutes of silent reading with pens out, then discussion anchored to specific page/line annotations.

When --silent-read is enabled, after generating the document the skill simulates the silent read by producing a second pass:

For each section, generate 2-4 questions a senior reviewer would write in the margin. Examples:

• "How confident is the 23pct number? Is this measured or estimated?"
• "Tenet 2 contradicts Tenet 4 — which wins when they conflict?"
• "Lesson Learned #3 says the prior attempt failed because of X. What evidence?"
• "Strategic Priority #1 has no owner named."

This catches the "writer thinks it's clear; reader doesn't" failure mode before the actual meeting.

Quality Bar

A six-pager passes if every check is true. Otherwise iterate.

• Page count ≤ 6 (memo) or correct sub-document length (prfaq)
• Every Goal has a number or specific verifiable outcome
• Every Tenet is a declarative commitment with no qualifiers
• State of Business cites at least 5 specific numbers / named systems / dated facts per page
• Strategic Priorities name at least one owner per priority
• Strunk audit: zero critical findings (passive voice, vague Goals, hedged Tenets); important findings acceptable with rationale
• The "what would cause us to kill this?" question (PRFAQ Internal FAQ) has a real, specific answer — not "if performance suffers"

Gotchas

• Do not exceed 6 pages. The constraint IS the value. If it doesn't fit, the thinking isn't done.
• Do not pad State of Business with backstory. Current data only. History belongs in Lessons Learned.
• Do not hedge Tenets. A Tenet with "try to" or "strive to" is not a Tenet.
• Do not generate fake numbers. If the Goal needs a metric and the user hasn't supplied one, ask. Generic placeholders ("X%") are not acceptable.
• Do not confuse Goals with Tenets. Goals are outcomes (measurable). Tenets are principles (commitments). Both are needed; conflating them muddles thinking.
• Do not skip the audit. Phase 4 is not optional. The whole point is prose discipline plus structure.
• Do not auto-save. Phase 6 confirms with the user where to write.
• Do not invoke this for prose articles. Use a dedicated long-form writing/voice skill instead — this skill is for decision docs, not essays.

Changelog

V2 (2026-05-27)

Optimized via skillforge optimize (outcome research: decision-memo efficacy + decision science). six-pager was already strong; this is a focused, non-duplicative add:

• Required premortem in memo mode — forward-looking "assume it failed, why?" failure analysis, each cause tied to a verification signal. Closes the gap that "Lessons Learned" (backward-looking) leaves open. Kahneman's highest-value decision technique; same decision-science line that informed council-review V2.
• Outcome target: memos that surface the killer risk before the decision, not after. Sources: Kahneman (premortem); Amazon narrative practice (silent read, customer-back); grounded in the same decision-science line (premortem, mediating assessments).

References

Bezos / Amazon canon:

• Bryar, Colin & Carr, Bill — Working Backwards: Insights, Stories, and Secrets from Inside Amazon (St. Martin's Press, 2021) — canonical book
• Slab — How Jeff Bezos Turned Narrative into Amazon's Competitive Advantage
• Maestra — How to Write an Amazon 6-Pager (with Template)
• Anecdote — What might Amazon's 6 page narrative structure look like?

Strunk & White canon:

• Strunk, William Jr. & White, E.B. — The Elements of Style, 4th edition (Pearson Allyn & Bacon, 2000)
• Public-domain text of Strunk's 1918 original: Bartleby — The Elements of Style

Anthropic prose-discipline:

• Anthropic — Claude Code Best Practices — the removability test

Cognitive-style background:

• Tufte, Edward — The Cognitive Style of PowerPoint (2003) — the case against bullet-points

Testing

This is a prompt-only skill. The shipped tests/eval.sh asserts the structural contract. To verify behavior end-to-end:

1. Pick a real decision: "/six-pager Should we migrate the API to cursor-based pagination?" (memo mode)
2. Verify all 6 sections generated, page count ≤ 6, audit run with findings
3. Pick a launch idea: "/six-pager prfaq Acme AI assistant public launch" (PRFAQ mode)
4. Verify press release + 2 FAQs generated, "what would cause us to kill this?" has a real answer
5. Run --strunk-only on an existing draft: verify only the audit fires, no new structure generated
6. Run --silent-read: verify margin questions are produced for each section