Writing Guidelines

Review the files or patterns provided by the user for compliance.

Read files, check against rules below. Output concise but comprehensive: sacrifice grammar for brevity. High signal-to-noise.

Rules

Planning & content type

• Every page has a plan (overview, goal, audience, content plan, open questions) referenced or linked
• Content type declared in meta.contentType: Tutorial, How-to, Reference, Conceptual, Troubleshooting, or Landing
• Title is user-shaped (the user's question), not feature-shaped (the engineer's name)
• Page does one job: tutorial OR how-to OR reference, not three at once
• Goal is verb-driven (Bloom's taxonomy): "configure", "explain", "debug" (testable)
• Multi-audience pages: short shared opener, then technical subsections

Voice & tone

• Active voice. Mental test: append "by monkeys". If the sentence parses, rewrite
• Direct address: you, never the user or one can
• Imperative for steps: "Click Add Project", not "You will need to click Add Project"
• Sentences under 20 words target
• Contractions encouraged (you'll, it's) for warmth
• Present tense unless describing future behavior
• Limit we: only for deliberate organization actions ("we recommend", "we deprecated"), never as a stand-in for "you"
• No rhetorical questions (sounds like marketing)
• Second-read test: read each sentence once at speech pace; if you re-read to parse it, name the subject, the action, and the consequence (kill metaphor verbs and pronouns reaching back several sentences)

Banned words

• easy, simple, quick: puts pressure on the reader and reads as marketing; replace with concrete description ("one command", "default settings", "most projects don't need this")
• very, just, really: filler; cut or rewrite

Concision

• Earn every detail: cut a number, name, or implementation detail if a more general phrasing wouldn't change the reader's understanding or action
• Weasel words: replace vague qualifiers (significantly, many, often, typically, generally) with a specific number or claim
• Vague quantifiers: no near-zero, sub-second, most requests; give the figure and cite it (99.37% of requests see zero cold starts)
• Filler/metaphor verbs: name the action instead of reaching for cadence (moves through, lands, carries, hits → the literal step)

AI-generated tells (flag these)

• Summary-style transitions: never open a paragraph by recapping the last one (With this setup complete…, Now that we've explored…); pivot straight to the next point (In practice…, The catch is…)
• Stop-start sentences: don't split one dependent idea into choppy fragments (Previously this was manual. Now it's automatic. This saves time. → one sentence); short sentences for emphasis are fine
• Spec-sheet voice: rewrite sentences that read like a system reading a datasheet (provides, is configurable, is explicitly labeled)
• Cold-open paragraphs: a body paragraph whose first sentence works as a standalone heading has no antecedent; carry the prior subject forward (Because…, Once…)
• Personified artifacts: machines don't perform human-physical actions (hand the browser a URL → the browser fetches the URL; the token holds… → the token is stored…)
• Reused framing: the angle must come from this page, not a template (The question most teams face is whether…)

Tone, by content type

• Tutorial: warm, encouraging, predictable structure, no traps
• How-to: terse, direct (reader is mid-task)
• Reference: neutral, exhaustive, quotable
• Conceptual: explain like the reader will teach it back; examples and analogies welcome
• Troubleshooting: empathetic but not apologetic; acknowledge then fix

Headings

• Sentence case for page headings (H1 H2 H3): "Configure environment variables", not "Configure Environment Variables"
• Title case for nav labels: "Configuring Environment Variables"
• meta.title becomes the H1; meta.navLabel becomes the sidebar entry
• Subheadings descriptive, not cute: "Caveats when self-hosting on Cloudflare", not "Caveats"
• Reader should be able to guess section content from the heading alone

Structure

• Every page opens with a one-paragraph TL;DR of what the page covers
• Every major section opens with a summary sentence
• Acronyms spelled out on first use: "Content Security Policy (CSP) blocks inline scripts"
• Define every term the first time you use it (link to its conceptual page)
• Reference docs organized by surface; education docs organized by reader task
• Keep paragraphs to 2 to 4 sentences; split anything longer or covering two ideas

Lists

• Three or more list-shaped items in a paragraph: convert to a list
• Bulleted for unordered; numbered for ordered (lifecycles, sequential steps)
• Always introduce a list with a colon
• No periods at the end of list items unless they are full sentences
• Bold/description format: - **Term**: description here (colon after bold term)

Code

• Code blocks need a language tag for syntax highlighting
• TypeScript is the default for new code unless the surface is genuinely language-agnostic
• Multi-step flows wrapped in <Steps/> so structure is visible
• Highlight load-bearing lines: ```typescript {8-12,23-37}
• ≤80 columns per line in snippets
• ≤25 lines per snippet; split longer blocks with prose
• Omit defaults; don't repeat variable definitions, use shared var
• Minimal comments in code blocks; prefer prose explanation
• Explain what every code block does in prose (don't drop and run)
• Don't reference full example files at the end of guides ("See train.py"); the guide is the deliverable

Placeholders

• Text placeholders: snake_case, descriptive: your_access_token_here (so reader can double-click to select before pasting)
• Number placeholders: count up 1234567890123 (recognizable as fake, predictable)
• Never <TOKEN>, xxx, your-token, or generic ALL_CAPS

Data sizes & units

• Space + uppercase unit: 64 KB, 5 KB, 200 ms
• Exception: seconds is bare: 30s
• Consistent across the corpus so readers can develop scanning habits

Money & pricing pages

• Uncompromising detail: err on "too much"
• Use tables for pricing
• Never assume reader knows the pricing model or whether their workload counts as one invocation or several
• Clarity and transparency above all else

Emphasis

• Bold means UI element or critical fact, never emphasis-for-emphasis-sake
• Reaching for bold for tone: the sentence is weak; rewrite it
• Inline code for paths, file extensions, identifiers, short snippets: /api, .tsx, body, query, req
• Rule: if it would look weird without a monospace font, monospace it

Punctuation & typography

• Never em dashes (—) or dashes (-) as punctuation; use colons, commas, periods, or rephrase
• Curly quotes " " and ' ', not straight " or '
• Ellipsis …, not three dots ...
• Loading states end with …: Loading…, Saving…
• Non-breaking spaces in 10 MB, ⌘ K, brand names
• & over "and" only where space-constrained (nav labels, buttons)

Source formatting

• Don't hard-wrap paragraphs: each paragraph is one line in source, let the editor wrap
• One blank line before headings; one blank line before and after code blocks
• No --- horizontal rules between sections
• No extra blank lines between elements that aren't paragraph breaks

Links

• Define every term the first time it appears, link to its conceptual page
• Anchor text names the destination; never bare URLs or here/link
• Dashboard links should open the exact destination and preserve required context in the URL.
• Link to canonical product documentation when relevant.
• Link to the canonical model catalog for the selected provider when examples name models.

Models in examples

• Always use the latest model strings: anthropic/claude-opus-4-7, not anthropic/claude-sonnet-4 or older
• For image generation default: google/gemini-3.1-flash-image-preview

AI workflow

• You are accountable for the content you produce, however it is created
• You are the final arbiter; the model proposes, you dispose
• Hold technical accuracy to a high standard: docs are also consumed by LLMs, wrong docs train wrong models
• Use only enterprise models that do not train on your data (especially for unreleased products)
• Disclose AI use in the PR (model + prompts if useful)
• Plan first by hand; the plan is the spec the model works against
• Use plan-mode in your editor (Cursor, Claude) before letting the model write
• Tell the model to follow AGENTS.md and the linting checklist
• Run a test prompt against the preview: "given this plan's goal, can the model complete the task using only this page?"
• Final human review always

Quality checklist (required boxes are non-negotiable)

• Findability: sidebar bucket set via meta.category; UI links to docs from any dashboard surface that exposes the feature
• Accuracy: code samples actually run; screenshots map 1:1 to current UI and use the ACME demo account
• Relevance: code samples included where applicable (TypeScript first; <Steps/> for multi-step flows)
• Clarity: overview addresses who/what/where/why; high-level use cases laid out; quickstart for new products; prerequisites listed on tutorials; a sample repository for multi-step tutorials; steps detailed not vague; visual aids in confusing sections; simplest path recommended when multiple exist
• Completeness: limits documented; all-limits tables updated; content plan followed and goals addressed
• Readability: nav names scannable and use action verbs; content types accurately used; subheadings descriptive; topics start with summaries; code blocks formatted correctly; active voice where warranted

Review

• PR description links to the content plan, lists what to review, and links the preview URL
• Ping the team via the PR link (not the plan or preview directly)
• Author is accountable, not the reviewer; reviewers are liberal with approvals
• Suggestion comments for small text fixes; preview comments for anything bigger
• Disagreement is fine; reject with a one-line reason and move on

Anti-patterns (flag these)

• Em dashes (—) or dashes (-) used as punctuation
• easy, simple, quick describing reader actions
• Passive voice (apply "by monkeys" test)
• Title Case in page headings (only sentence case in H1 through H6)
• Generic placeholders: <TOKEN>, xxx, your-token, ABC123
• Code blocks without a language tag
• JS examples where TypeScript is the convention
• Code blocks over 25 lines without prose between
• Hard-wrapped prose paragraphs (multiple lines for one paragraph in source)
• --- horizontal rules between sections
• Subheadings that are single generic words: Overview, Caveats, Notes
• Bold used for emphasis instead of UI element or critical fact
• Page or section without an opening summary
• Straight quotes (", ') instead of curly (", ')
• Three dots (...) instead of ellipsis (…)
• Acronyms used before being spelled out
• Bare unit numbers (64KB, 5kb, 200MS) instead of 64 KB, 5 KB, 200 ms
• "We" standing in for "you"
• Rhetorical questions
• Filler words: very, just, really, simply
• References to "the full example file at the end of the guide" rather than inlining the code
• Outdated model strings in examples (anthropic/claude-sonnet-4, gpt-4o, DALL-E)
• Hardcoded date/number formats instead of Intl.DateTimeFormat / Intl.NumberFormat in code samples
• "Loading..." instead of "Loading…"
• Summary-style transitions recapping the previous paragraph (With this setup complete…)
• Stop-start fragments splitting one dependent idea into choppy sentences
• Spec-sheet voice reading like a datasheet (provides, is configurable, is explicitly labeled)
• Cold-open body paragraphs whose first sentence has no antecedent
• Personified artifacts performing human-physical actions (hand the browser a URL)
• Reused/template framing not specific to the page (The question most teams face is whether…)
• Weasel words instead of a specific claim (significantly, many, often, typically, generally)
• Vague quantifiers without a cited figure (near-zero, sub-second, most requests)
• Filler/metaphor verbs instead of the literal step (moves through, lands, carries, hits)
• Sentences that need a second read to parse
• Paragraphs over 4 sentences or covering two ideas
• Bare URLs or here/link as anchor text

Output Format

Group by file. Use file:line format (VS Code clickable). Terse findings.

## content/docs/sandbox.mdx

content/docs/sandbox.mdx:1 - missing meta.contentType
content/docs/sandbox.mdx:12 - title "Cloud Sandbox" is feature-shaped, not user-question
content/docs/sandbox.mdx:24 - passive voice ("the sandbox is created...")
content/docs/sandbox.mdx:31 - banned word "easy"
content/docs/sandbox.mdx:47 - "..." → "…"
content/docs/sandbox.mdx:58 - code block missing language tag
content/docs/sandbox.mdx:71 - placeholder <TOKEN> → your_access_token_here
content/docs/sandbox.mdx:89 - "64KB" → "64 KB"
content/docs/sandbox.mdx:102 - H2 "Caveats" too generic; add specificity
content/docs/sandbox.mdx:118 - em dash in prose, replace with colon/comma

## content/docs/ai-gateway.mdx

content/docs/ai-gateway.mdx:5 - title case in H1; sentence case only
content/docs/ai-gateway.mdx:18 - acronym AI Gateway used before being spelled out
content/docs/ai-gateway.mdx:34 - bold for emphasis, not UI element
content/docs/ai-gateway.mdx:52 - `anthropic/claude-sonnet-4` outdated; use `anthropic/claude-opus-4-7`
content/docs/ai-gateway.mdx:71 - hard-wrapped paragraph (lines 71-74)

## content/docs/cron.mdx

✓ pass

State issue + location. Skip explanation unless fix is non-obvious. No preamble.