Field Notes
The notebook, not the write-up. Subjectless, present tense, and mostly not in sentences.
The feeling Being in the middle of it, before you know what it means.
Delete the subject. The notebook runs first-person pronouns at 87 per 10,000 words. Darwin's own write-up of the same days runs 242, and his published book 177. The "I" is what retrospection adds, because you only narrate yourself once you have stepped back far enough to watch yourself doing it.
/stet voice field-notes
Open the rules
Rules
- Present tense for the thing in front of you, bare past for what you did
- The doubt is a question mark, not a hedge
- Numbers, at eight to eleven times the write-up's rate
- The colon abuts, it does not argue
- The admin stays in the book
- The date is a header, not a sentence
- One adjective, then stop
Never
- Never second person. 9 tokens in 110,671 words, most of them reported speech.
- Never state a relation between two observations. The colon abuts. The argument comes later or not at all.
- Never close the line. 84 percent of units do not end in terminal punctuation.
- Never frame retrospectively. No "as it turned out", no "in the event", no summary of the day. Summary is the diary's job and the diary does it 1,661 times.
- Never write a paragraph. The notebook's unit median is 10 words. It has lines, not paragraphs.
- Never tidy out the admin.
- Never round a number.
How pastiche fails
- Pronoun inflation, about three times.
- Hedge inflation, about 2.5 times, on the wrong instrument.
- Second person, 40 to 50 times over.
- Variance collapse.
- Terminal punctuation.
- Numeral starvation.
- Standard deviation over mean.
- First-person pronouns per 100 words.
- Share of lines ending in a full stop.
- Numerals per 100 words.
Measured from Darwin's 15 Beagle field notebooks, Darwin Online; Darwin, Beagle Diary F1925; Darwin, Voyage of the Beagle; Lewis and Clark journals; Scott, Last Expedition vol I; Thoreau, Journal I
Read the whole file
Patch Notes
A changelog. One change per line, verb first, numbers exact, and the joke under five percent.
The feeling Respect, mostly. Occasionally delight, and never at somebody's expense.
The utility carries the voice, never the other way round. A joke reads as funny only against a wall of flat, useful lines. Raise the density and the reader stops trusting any line to be literal, which is the failure this register dies of.
/stet voice patch-notes
Open the rules
Rules
- One change per line
- One grammatical shape per line type
- Numbers exactly, in one convention
- The direction verb carries the nerf
- The order never changes
- Deadpan beats a joke
Never
- Never joke on a security, data-loss or breaking-change entry. The funny products are silent exactly where the stakes are highest.
- Never joke on a nerf. It reads as gloating at the player it affects.
- Never write a fake entry. It costs the reader a lookup, and when they find nothing they distrust every line beside it.
- Never explain the joke, and never signal it with an emoji or a wink.
- Never announce enthusiasm. No "we're excited to", no "thrilled", no exclamation marks outside a deliberate aside.
- Never use marketing adjectives on your own work: powerful, seamless, revamped, game-changing.
- Never hedge a number.
- Never put justification inside an entry.
- Never use first person for the fix. "Fixed X", not "we went in and fixed X".
- Never joke about the team. The real ones joke about the artifact or the absurdity of the defect, never about the developers' suffering or their coffee intake.
How pastiche fails
- The ratio inverts.
- The joke replaces the fact.
- Length inflation.
- In-jokes needing context the reader lacks.
- Puns on the product name.
- Applied where nothing changed.
Measured from Valve (Team Fortress 2, Counter-Strike 2), Blizzard (World of Warcraft, Overwatch), Supergiant (Hades), Slack, Apple, 37signals
Read the whole file
Plainspoken
Concrete nouns, working verbs, no throat-clearing. Varied hard, because uniform length is the tell.
The feeling Conviction, and the calm of somebody who does not need to convince you.
Vary hard. Variance is the register, not shortness.
/stet voice plainspoken
Open the rules
Rules
- Concrete over abstract
- Let the meaning choose the word
- Keep the hedge that carries a fact
- Do not chop at the conjunction
- No throat-clearing
- The short sentence is the verdict
Never
- Em dashes.
- "Not X, it is Y." Say Y.
- A summary of what the reader just read.
- Adjectives standing in as instructions.
- Explaining to readers what they obviously are, or why they are here.
- Restating a claim in different words to be sure it landed. See below: that is where condescension actually comes from.
How pastiche fails
- The variance collapse.
- One subject, one shape, three times.
- Mechanical de-Latinising
- Hedge-stripping
Detection standard deviation over mean, below 0.35. And three consecutive sentences opening with the same word, unless it is deliberate anaphora, which is distinguishable because deliberate anaphora runs three or more with escalation rather than two by accident.
Measured from Orwell, Hemingway, GOV.UK content guidance, digital.gov, Butterick, Paul Graham, and Pullum's criticism of Strunk and White
Read the whole file
The Argument
Essayistic. Wide variance, concession before the strongest claim, and an ending that never summarises.
The feeling The pleasure of watching somebody think, and change their mind on the page.
Make claims as strong as they can be made without becoming false. Graham's dial, and the whole register in a sentence. Weaker than that and the essay is not worth reading. Stronger and it is not worth trusting.
/stet voice the-argument
Open the rules
Rules
- Variance carries the argument
- Keep the qualifier, and pick the right one
- Concede before the strongest claim, never after
- The thesis may move, and that movement is the essay
- Move downward, from the general to the particular
- Brevity is the diction of command
Never
- Never restate the argument at the end. The conclusion-that-recaps is a vestige of legal rhetoric, where it was the closing remarks to a jury, surviving in a form whose purpose is not persuasion.
- Never peter out. The failure Graham names is starting on the most radioactive question and then recoiling from it, ending on "one thing is certain: the question is a complex one."
- Never ask a rhetorical question you do not then answer with a fact.
- Never say "clearly" or "obviously". If it were, the sentence would be unnecessary.
- Never build a strawman. It costs more credibility than it wins.
- Never digress without paying for it. Digression is licensed, but under a constraint: the essay owes the reader something they did not already know.
How pastiche fails
- claim-altitude with no evidence under it
- The aphorism with no referent.
- The unearned "we".
- The colon that promises depth
- "Not X but Y", saturated.
- The escalating tricolon.
- The summary ending in disguise.
Detection count the sentences making a general claim with no number, name, date or object in them. In the measured essays that ratio is low and clusters at section ends. Uniform across the text, the piece is intoning rather than arguing. Then, for each general sentence, ask what specific observation produced it. If there is none, it is decoration.
Measured from Orwell, Didion, Baldwin, Paul Graham, McPhee, Pullum, Zadie Smith, Hitchens
Read the whole file
The Broadsheet
Wire news. Most important fact first, every claim sourced, and "said" is the only verb.
The feeling None, on purpose. Know that before choosing it.
Every claim is sourced, and the sourcing verb is "said". Not "according to", not "revealed", not "noted". Reuters: those verbs "all contain an element of judgment by the reporter."
/stet voice the-broadsheet
Open the rules
Rules
- Count the words in the first sentence
- News first, source second, unless it is contentious
- Two adjectives, then rewrite
- Days of the week, never yesterday
- Give the actor
- Never adopt the subject's framing
- Numbers, and repeat the surprising ones
Never
- Never put opinion in a news story. Features get no exemption.
- Never leave a claim unattributed. Contentious statements are re-sourced every time they appear, and "alleged" is not a substitute for a source.
- Never use a judging adjective or adverb: a hard-line speech, a glowing tribute, a staunch conservative.
- Never use an unverifiable superlative: unique, unprecedented, first, only, largest, record. If somebody else claims it, attribute the claim to them.
- Never read minds. "To write that someone hinted, implied, indicated, suggested, or signaled is to interpret someone's thoughts. This is rarely acceptable."
- Never qualify a denial. No "flatly denied", no "categorically denied", unless it is inside a quote.
- Never use journalese: reportedly, probe, dubbed, slammed, massive, embattled, oil-rich, looks set to, hit by fears that.
- Never use euphemism. "Died of cancer", not "passed away after a fight with cancer".
- Never alter a quotation, even to fix grammar.
- Never boast about your own reporting.
- Never begin with a question. "The audience wants to be informed, not take part in a quiz."
How pastiche fails
- Journalese standing in for reporting.
- The sympathetic attribution verb.
- The unattributed background paragraph.
- The lede that summarises rather than reports.
- Adopting the subject's own name for the thing.
- The unverifiable superlative, unattributed.
- Mind reading.
- The pyramid shape with an argument inside it.
Detection count the adjectives and adverbs that could not be checked by a third party, then count the sentences with no attribution. In wire copy the first number is near zero and the second is near zero. Either one rising is the writer becoming visible.
Measured from Reuters Handbook of Journalism, AP Statement of News Values, BBC News Styleguide, The Economist Style Guide
Read the whole file
The Catalogue
The object record. Fields rather than sentences, and the same words on purpose.
The feeling Authority, manufactured by the format rather than earned by the words.
The same words, on purpose. Between 23 and 50 percent of the sentences in these documents are verbatim repeats of another sentence in the same document. The commonest way to begin a lot in the 1898 catalogue is the single word "Another.", which opens 69 of its 301 lots.
/stet voice the-catalogue
Open the rules
Rules
- Fields, not sentences
- The adjective and the measurement both, in separate slots
- Use the worn-out evaluative
- Field order fixed, field contents free
- Uncertainty occupies the slot rather than emptying it
- Close on the measurement
- Leave the gaps
- Discussion goes somewhere else, physically
Never
- Never a question. 0 in 14,218. Also 0 across 3,433 interpretive notes.
- Never an exclamation. 1 in 14,218. The commercial layer breaks this 21 times, and the Whole Earth Catalog runs 20.1 per 10,000 words, so it is a rule about object records alone.
- Never second person. 0.00 here, and untrue of anything commercial.
- Never open on an article. Between 0 and 9 percent of entries begin with "The", "A" or "An".
- Never a connective. Between 0.0 and 0.4 per 10,000 words in the tightest corpora. The auction catalogue at 16.0 and Burpee at 26.9 are not obeying this, and both are looser registers.
- Never a hedge about your own confidence. 0.01 per sentence.
- Never a paragraph inside an entry. Two rules that were on this list and are measured off it: never repeat, and never evaluate. Both are
inversions, and both are in the rules above because they are what the register actually does.
How pastiche fails
- It writes sentences, so the field count collapses.
- It is too smooth.
- It never repeats itself.
- It avoids "very".
- It picks one layer and runs it all page.
- It fills every field.
Measured from American Antiquities auction catalogue 1898; Rock, Textile Fabrics, South Kensington Museum 1870; Catalogue of the Gallery of Art, New York Historical Society 1915; Cleveland Museum of Art open access records; V&A physicalDescription; Sears Roebuck Consumers Guide 1897; Burpee's Farm Annual 1885; Whole Earth Catalog 1968
Read the whole file
The Manual
Reference writing. It describes behaviour, and only the steps give orders.
The feeling None, on purpose, and for a reader who is already frustrated.
Complete beats readable. A reference page is not read, it is consulted, by somebody who already has a problem. Leaving out an edge case to keep a paragraph tidy fails the one person who came looking for that edge case.
/stet voice the-manual
Open the rules
Rules
- Condition before instruction
- Qualify with a modal, never with an adverb
- Put the edge case in brackets
- State the default with its alternatives
- Document the precedence
- Send the reader somewhere
- Open a task page by saying what it does
- Contractions are fine
Never
- Never a question. 0.003 of sentences, and zero across the Python and MDN samples.
- Never an exclamation mark. 39 in 385,405 words.
- Never an em dash. 136 in the corpus, and zero in both the GNU coreutils manual and the man page option glosses. Procedure documentation runs one em dash in 161,019 words.
- Never first person singular. 0.9 per 10,000, and nearly all of that is "I/O".
- Never the marketing verbs. "leverage", "robust" and "seamless" are at 0.0 per 10,000 in reference.
- Never past tense for current behaviour. Where reference uses it, it is version history.
- Never a uniform sentence length. The standard deviation over mean is 0.73 with a tail to 321 words, the longest being in the GNU coreutils manual. One rule here is a correction rather than a description, and it is worth keeping while being honest
about that. Never "simply", "just" or "easy". Google bans them, and the practice does not obey:
they run at 4.9 per 10,000 across the corpus, 11.1 in the GNU coreutils manual, and
open(2) uses
"simply". They tell a reader who is stuck that their problem is beneath notice, which is the one
thing this register exists not to do.
How pastiche fails
- Second person inflation, by three to thirteen times.
- Tail flattening.
- Hedge substitution.
- Cross-reference suppression.
- The missing conditional.
- Parenthesis avoidance.
Measured from Linux man-pages (18 pages); GNU coreutils manual 9.11; Python 3 library reference (10 modules); MDN (15 pages); Kubernetes task documentation (93 files); Raspberry Pi documentation (78 files); GitHub Docs (72 files); Google developer documentation style guide, CC BY 4.0 (15 pages)
Read the whole file
The Teacher
Explains. Names the confusion before it happens, and bounds every analogy before the reader over-extends it.
The feeling Warmth, and the specific relief of being taken seriously.
Describe the material, never the reader's mind. This is the whole diagnostic for condescension, and it fits in a sentence.
/stet voice the-teacher
Open the rules
Rules
- Long sentences, varied hard
- Three pronouns, three jobs
- The name arrives after the thing
- Name the confusion as a fact, not a question
- Predict the wrong inference, then cancel it
- Every analogy carries its own expiry
- Reassurance must be attached to a fact
- Say where every claim stands
- Retract the scope at the end
Never
- "Just". A filler that presumes a background. Usually deletable with no loss of meaning.
- "Simply", "easy", "trivially", "of course", "everyone knows", "as you'd expect". All share one mechanic: they assert the reader's mental state, so a reader who is stuck is told in passing that being stuck is anomalous.
- "Obviously" and "clearly" aimed at the reader. One licensed exception, from Halmos: you may call something obvious to place it in perspective, and if you do, make sure the obvious thing is true. Nielsen's only real use of it is aimed at his own model. Point it at your own claim, never at the reader's comprehension.
- Tag questions. "Makes sense?", "See?", "Pretty simple, right?" Each demands assent and offers no way to withhold it.
- A rhetorical question you answer yourself in the next clause. It stages a dialogue the reader is not in.
- Bluffing. Halmos: readers sense concealment, and they blame neither the facts nor themselves. They blame the author, correctly.
- An unbounded analogy, or one quietly doing a second job it was not introduced for.
- Simplifying the vocabulary below the reader's level. It buys nothing, measurably.
How pastiche fails
- no more accurate
- The collective pronoun misused.
- Reassurance with no content.
- Uniform short sentences.
- Enthusiasm standing in for structure.
- The analogy that is never taken away.
- Warm tone, expert sequencing.
Measured from Feynman, Sagan, Bartosz Ciechanowski, Michael Nielsen, Grant Sanderson, Bret Victor, Paul Halmos, Julia Galef
Read the whole file