Structure readers can scan

The choice between list types is not aesthetic. It is a signal to the reader about whether they can rearrange the items:

• Bulleted list: order does not matter. Rearranging the items changes nothing. Valid event types, prerequisites, known limitations - all bulleted.
• Numbered list: order matters. Rearranging the items breaks the meaning. Pulse's run order (ingest, then clean, then metrics) must be numbered - a bulleted version silently tells readers they can run the scripts in any order.
• Start each numbered step with an imperative verb. "Run the ingest script", not "The ingest script should be run" and not "Running the ingest script". Command verbs make steps executable at a glance.

And parallelism, the rule that makes lists feel professional: every item in a list matches every other item in grammar, category, capitalization, and punctuation. One item that starts with a noun in a list of verbs reads like a wrong note in a chord.

Google highlights this rule because getting it backwards causes real damage: state the condition BEFORE the instruction, not after. Readers execute instructions the moment they read them - a condition that arrives afterwards arrives too late.

In pipeline docs this pattern is everywhere: "If metrics.py reports LOW days, check alerts.min_daily_events before paging anyone." The reader who needs the caveat gets it before they act; the reader who does not skips the whole line from the first two words.

A list holds one dimension of information. The moment each item carries two or more facts - a setting AND what it does AND whether you can touch it - a list forces the reader to parse prose for structure. A table hands them the structure directly:

Table hygiene, per Google:

• Label every column with a meaningful header - readers scan headers before cells.
• Keep cells short. If a cell needs three sentences, the content wants to be a paragraph or its own section, not a cell.
• Introduce the table with a sentence ending in a colon, preferably containing "following": "The following table describes every config.yaml setting:"

Google's paragraph unit has one non-negotiable rule: the opening sentence establishes what the whole paragraph is about. Scanners read lead sentences and nothing else, so the lead sentence IS the paragraph for most readers.

• The test: read ONLY the first sentence of every paragraph in your doc, top to bottom. If the story survives - you can still follow what the doc says and why - your structure works. If the story falls apart, your key points are buried mid-paragraph where scanners never look.
• One topic per paragraph. The moment a second topic sneaks in, split. A paragraph is a unit of thought, not a unit of typing.
• Three to five sentences. Under three, consider merging or promoting to a list. Over seven is a wall - readers' eyes slide off it, and walls are where facts go to hide.

Headings are the largest text on the page and the first thing scanned, yet most docs treat them as an afterthought. Microsoft's scannable-content guidance is specific:

• Sentence case, always: "Run the pipeline", not "Run The Pipeline". Google's style guide agrees, so this one is settled across the industry.
• No end punctuation on headings. No periods, no colons.
• Task-based over topic-based. "Configure alert thresholds" tells the reader what they will be able to do; "Alerts" makes them open the section to find out. Verb-first headings double as a table of contents of actions.
• Short. A heading that wraps to two lines is a sentence wearing a costume.

Google's completeness test for a paragraph: a good paragraph answers three questions. Not always in this order, but all three should be findable:

• What are you telling the reader? The core fact or instruction.
• Why does it matter to them? The consequence of knowing or ignoring it.
• How should they use the knowledge? The action it enables, or the way to verify it worked.

Run the check on a paragraph like "Pulse flags LOW days." What: any day with fewer events than alerts.min_daily_events is flagged. Why: a LOW day usually means an upstream export failed, not a real traffic drop. How: check the raw folder for that date before trusting the metric. Three sentences, three answers, one complete paragraph.