One skill for Amazon narrative document types — shared spec format for humans reviewing in meetings and agents executing in spec-driven development.
Companion wedge: four-answers — answer-only enforcement on Q&A.
Deep guides: docs/ · Rules reference: reference.md · Spec-driven context: docs/spec-driven.md
Why this matters now (spec-driven development)
In spec-driven development, the document is the contract:
- Humans read it in narrative meetings (silent read → FUND / KILL / REVISE)
- Agents read it as the source of truth before writing code, tests, or implementation plans
- Both need the same properties: precise language, testable claims, explicit risks, no weasel words
A PR/FAQ is not just a product ritual — it is a structured spec that an agent can:
- Parse into requirements and non-goals
- Lint for ambiguity before implementation starts
- Trace decisions back to customer problem statements
- Fail fast when internal FAQ questions lack four-answer form
Amazon's formats predate AI agents, but they solve the same problem as modern spec.md / AGENTS.md workflows: force clarity before execution.
When drafting for agents, add:
- Explicit in scope / out of scope
- Acceptance signals — how humans and agents know the spec is satisfied
- Kill criteria — when to stop building
- File paths or artifacts agents should produce (when user provides a repo)
Universal rules (every doc type, human + agent)
Apply on draft and review:
1. Fewer than 30 words per sentence
One idea per sentence. Agents parse long sentences into wrong requirements.
| Bad (43 words) | Good |
|---|---|
| Due to the fact that customers were experiencing significant delays, we made the system much faster. | We cut server-side p90 latency from 10 ms to 1 ms. |
2. Replace adjectives with data
Subjective adjectives are untestable for agents and misleading for humans.
| Bad | Good |
|---|---|
| significantly better conversion | conversion +2.3 pp (4.1% → 6.4%) |
| most customers | 87% of Prime members |
| much faster | p90 latency 10 ms → 1 ms |
3. Eliminate weasel words
Weasel words let agents hallucinate compliance and let humans nod without deciding.
Banned in conclusions: maybe, perhaps, probably, likely, potentially, could, might, seems, appears, I think, I believe, arguably, generally, usually, nearly all, significantly, substantially, would help, might bring clarity.
Exception: quoting a third party's hedged statement — label it as their view.
4. "So what?" test
Before every paragraph: if deleted, would the reader (human or agent) miss it? If no, cut it.
5. Four answers
On substantive questions: Yes, No, a number, or I don't know but I'll know it by X.
Bare "I don't know" is invalid. Agents must either have evidence or a named artifact/date that will produce evidence.
Pick a document type
| User wants | Doc type | Read |
|---|---|---|
| Fast decision, go/no-go | 1-pager | docs/1-pager.md |
| Deep narrative, silent-read meeting | 6-pager | docs/6-pager.md |
| New product / agent feature spec | PR/FAQ | docs/pr-faq.md |
| Durable decision principles | Tenets | docs/tenets.md |
| Post-incident / failed spec | COE | docs/coe.md |
| Agent + human spec workflow | Meta | docs/spec-driven.md |
| Review existing doc | Lint | Rules above + doc-type checklist |
If unspecified, ask once: decision (1-pager), deep dive (6-pager), new product/spec (PR/FAQ), principles (tenets), incident (COE), or lint?
Workflow
Draft (human or agent-authored)
- Confirm doc type
- Read the full matching
docs/*.md— do not improvise structure from memory - Gather evidence: customer problem, metrics, competitive context, risks
- Draft in full prose — not bullet slides with periods
- Add agent-facing blocks when spec-driven: scope, non-goals, acceptance signals
- Run lint pass before delivering
Review / lint (human or agent reviewer)
Output this format:
AMAZONIAN LINT
Doc type: PR/FAQ
Audience: human leadership + implementation agent
Score: 72/100
BLOCKERS (must fix before FUND or agent execution)
- Internal FAQ Q3: "probably" → number or IDK BY [date]
- Problem paragraph: "significantly better" → +25 bps or delete
- No explicit out-of-scope section for agent
FIXES (should fix)
- Sentence 4 (42 words): split
- Missing kill criteria in internal FAQ
AGENT READINESS
- [ ] Requirements traceable to problem paragraph
- [ ] Non-goals stated
- [ ] Testable acceptance signals present
- [ ] Risks have owners and dates
Blockers = weasel words, hedged answers, untestable claims, missing four-answer form.
Fixes = sentence length, weak data, "so what?" failures.
Meeting / spec gate decision
End with exactly one:
- FUND — proceed to build / agent may implement
- KILL — do not build; state why
- REVISE — named owner + return date + what evidence is missing
When this skill applies
Apply when:
- Drafting or reviewing 1-pager, 6-pager, PR/FAQ, tenets, COE
- User says "Amazon memo," "narrative," "Working Backwards," "Write Like an Amazonian"
- Spec-driven: PR/FAQ as feature spec, 1-pager as decision gate before agent work
- Linting docs before human review or before handing spec to coding agent
Do not apply when:
- Pure code implementation with an approved spec already in context
- Creative writing unrelated to business decisions
- Quick Yes/No only → four-answers
Doc type index (summaries — read full docs before drafting)
| Type | Max size | Primary audience | Agent use |
|---|---|---|---|
| 1-pager | ~1 page | Decision maker | Decision gate before sprint |
| 6-pager | 6 pages | Leadership meeting | Strategy context for multi-sprint work |
| PR/FAQ | 6 pages total | Customer + leadership | Feature/product spec |
| Tenets | 5–10 rules | Team | Constraint layer for agent judgment |
| COE | 2–4 pages | Engineering + ops | Postmortem; update specs/mechanisms |
Full templates, examples, anti-patterns, and agent mapping in each docs/*.md file.
Final check
- Read the full doc-type guide for this format?
- Human-readable and agent-parseable?
- Every sentence <30 words?
- Data instead of adjectives?
- Zero weasel words in conclusions?
- Four-answer form on substantive claims?
- PR/FAQ: believable customer quote + internal FAQ names biggest risk?
- Spec-driven: scope, non-goals, acceptance signals, kill criteria?
- Ends with FUND / KILL / REVISE when appropriate?