Farnsworth Rhetoric: Memorable Writing

Apply classical figures to writing that already exists. This skill does not generate messages and does not restructure arguments. It works on the sentence and the paragraph — word choice, parallel structure, and where the emphasis falls.

Figures are expensive. Each one draws attention to itself, and attention spent on the sentence is attention not spent on the point. Most text needs one figure. Some needs none. The dosage budget below is a hard constraint, not a suggestion.

Workflow position

Run this skill last.

Long-form:   human-narrative → humanizer → farnsworth-rhetoric
Short-form:  humanizer → farnsworth-rhetoric

Figures applied before humanizer get stripped by humanizer. Figures applied to text that still carries AI tells amplify those tells — a rhetorical flourish on top of promotional prose reads worse than the promotional prose alone.

This skill's output is designed to pass humanizer. If you re-run humanizer afterward and it flags something, check "Forbidden constructions" below. If the construction isn't on that list, the figure is intentional — leave it.

How to use this skill

1. Read the whole piece before changing anything. Find the one sentence that carries it.
2. Triage: set the context, the register, and the figure budget (Step 1 below).
3. Diagnose flatness — where does the prose fail to land? Endings first; they matter most.
4. Check each candidate figure's trigger condition. No trigger, no figure.
5. Apply within budget. Work backward from the ending.
6. Run the claim check and the ear test. Both are mandatory.
7. Report what you changed and what you deliberately left alone.

When one line carries the whole piece — a tagline, a closing sentence, a headline — chain beyond-obvious to generate 3–5 distinct rhetorical treatments before choosing. Applying the first figure that fits is the most common failure mode of this skill.

Load references/figures.md for figures not covered here: anadiplosis, symploce, asyndeton and polysyndeton, epizeuxis, praeteritio, litotes, and strategic passive voice. Load it when the core figures below don't fit, or when the user names a figure by its classical name.

Use cases

Diagnose only: User asks "why does this fall flat?" or "where could this be sharper?" Report where the prose loses energy, name the figure each spot would take, and give the budget. Don't rewrite until asked.

Strengthen: User says "make this land" or "farnsworth this." Triage, apply within budget, return the revised text plus a change note naming each figure and the trigger that justified it.

Single line: User wants one sentence sharpened — a tagline, a closing line, a slide title. Chain beyond-obvious for options, present 3–5 treatments with the figure named, and recommend one. Budget is 1 figure. Run the claim check on every treatment before presenting it; a treatment that fails is withdrawn, not offered with a caveat. Output shape: the treatments list, each with its figure named, then one change note for the recommended treatment only.

Step 1: Triage

Answer three questions before touching the text.

Q1: What is the register?

Q2: What is the figure budget?

Count the words. Do not estimate them; the budget tiers have hard edges.

Additional hard caps:

• Never two figures from the same family in one paragraph.
• Never the same figure twice in one piece. One anaphora run counts as one figure, not three.
• Zero is a valid budget spend. Say so if the text doesn't need a figure.

Q3: What is free?

Diction work does not count against the budget, because it removes ornament rather than adding it. Apply it everywhere, always — but free of budget is not free of the claim check below: a Saxon-default swap can drop an attribution or a hedge as easily as a figure can.

• Saxon default. Trigger: three or more Latinate polysyllables in one sentence, or a chain of abstract nouns. Move: turn nominalizations back into verbs. "The implementation of our transformation initiative will require" → "We will have to." Corpus comparison finds Latinate polysyllables and nominalisations elevated in current model output, so this rule has empirical support and not only a craft argument.
• Saxon finish. Trigger: a sentence or paragraph ends on a Latinate polysyllable (-tion, -ment, -ity, -ance, -ize). Move: rebuild so the last word is a stressed Saxon monosyllable. The last word rings longest. Strong finishers: work, fight, stand, fall, hate, love, live, die, win, lose, end, start, break, build, cost, pay, lost, gone.

Holmes: "...freedom for the thought that we hate." Four Latinate-free words after a formal legal buildup. "Hate" is the punch, and it's the last thing you hear.

Master principle: contrast

The ear detects differences, not qualities. A Saxon word lands because Latinate words surround it. A short sentence hits because long ones preceded it. Every figure below is a way of engineering a difference.

This has a practical consequence: a figure with nothing to push against does nothing. Six punchy Saxon sentences in a row are as flat as six Latinate ones. If you cannot name what the figure contrasts with, don't apply it.

Core figures

Each figure has a trigger (the detectable condition that licenses it), a move, and a cap. Skip any figure whose trigger is absent. Do not go looking for triggers.

Antithesis — the figure of contrast

Trigger: You can name what the thing is not, and both poles are real.

Move: Put the two poles in parallel grammatical frames, adjacent, with equal weight.

Swap test: Reverse the poles. If the sentence still makes sense, the contrast is fake — you have two unrelated statements in a parallel costume. Cut it.

Franklin: "We must indeed all hang together, or most assuredly we shall all hang separately."

Before: "Our pricing is competitive and our support is a differentiator." After: "Our competitors win on price. We win after the sale."

Not the same as negative parallelism. "It's not just X, it's Y" asserts one pole and dismisses a straw pole — humanizer flags it (NEG-PARALLEL) because the dismissed pole was never a real option. Antithesis gives both poles weight and lets them stand.

Cap: 1 per three paragraphs.

Isocolon — parallel members of equal shape

Trigger: Two to four items that carry equal weight and can take equal grammatical shape. Three is the number that resolves; two feels open, four feels like a list.

Move: Match clause length and structure so the members scan alike.

Load-bearing test — required. An isocolon earns its place only if each member adds something the others don't. Ask: could I cut this to one member and lose information? If no, cut it to one. Interchangeable adjectives and padded parallel phrases are humanizer's decorative triad (RULE-OF-3), not a figure. humanizer states the same load-bearing test from its side, so the two skills agree on which triads survive.

Fails the test: "We value innovation, collaboration, integrity, excellence, and growth." (Five interchangeable abstractions. Cut to what you actually do.) Passes: "We came, we saw, we conquered." (Three distinct events, in sequence, each required.)

Cap: 1 per piece short-form, 1 per section long-form.

Antimetabole — words reversed

Trigger — test all three:

1. "They have it backwards." The relationship is inverted.
2. "And vice versa." The situation is genuinely reciprocal.
3. "It's the other way around." There's a mismatch to correct.

Move: Repeat the same words in reverse order. A-B-B-A, with word identity.

Kennedy: "Ask not what your country can do for you — ask what you can do for your country." (country → you → you → country)

Before: "Don't let work define your life. Make sure your life shapes your work." After: "Don't let your work define your life. Let your life define your work."

Antimetabole vs. chiasmus: Antimetabole reverses the same words and is therefore checkable — you can verify it mechanically. Chiasmus inverts parallel structure without word repetition and is much harder to land. Default to antimetabole. Attempt chiasmus only when the antimetabole is unavailable and the inversion is genuinely worth having.

Cap: 1 per piece, maximum. Most pieces should have none. This is the most over-attempted figure in the catalog — if the trigger tests don't fire cleanly, the result will read as a fortune cookie.

Hypophora — ask, then answer

Trigger: The reader has an objection or question you are about to address anyway.

Move: Ask it in their words. Answer it in the next sentence.

Why it works in business writing: It names the objection instead of routing around it, which is why it reads as confidence rather than ornament. It also turns the reader from evaluator into participant.

Before: "There are cost considerations associated with the migration timeline." After: "So why not wait a year? Because the license renews in March, and renewing locks us in for three more."

Anti-patterns: Never ask a question you don't answer immediately. Never announce it ("You might be wondering...") — humanizer flags that as SIGNPOSTING. Just ask.

Cap: 2 per piece. The most under-used figure in business writing; spend here first.

Erotema — the rhetorical question

Trigger: The answer is obvious, and the obviousness is the point.

Move: Ask. Don't answer.

Cap: 1 per piece, and prefer hypophora instead. In business writing an unanswered question often reads as evasion or as a rhetorical move the reader has seen too often. Use it when the audience already agrees and you want them to say so to themselves.

Anaphora — repeated opening

Trigger: Three or more parallel commitments or actions sharing one subject, at the emotional peak of the piece.

Move: Repeat the opening words verbatim across consecutive clauses.

Churchill: "We shall fight on the beaches, we shall fight on the landing grounds, we shall fight in the fields."

Cap: 1 run per piece, at the peak. Never in anything under 300 words. Anaphora in an executive summary or an email reads as speechifying, because it is. This is the figure most likely to blow the whole budget in one paragraph — the run is one figure, but a second run in the same piece is a tell, not a technique.

Epistrophe — repeated ending

Trigger: One word or phrase is the piece's actual subject and can fall naturally at the end of consecutive clauses.

Move: End consecutive clauses on it. The repeated word lands in the most memorable position.

Lincoln: "government of the people, by the people, for the people."

Cap: 1 run per piece. Do not combine with anaphora in the same paragraph — the two compete for the same attention, and the ear reads the collision as a stutter. Farnsworth's shift from one to the other across a longer passage (symploce) is covered in references/figures.md.

Forbidden constructions

These are figures the model reaches for by default, and each one is on humanizer's flag list. Producing them makes this skill actively counterproductive.

EM-DASH is stricter here than in humanizer, deliberately: this table governs what this skill produces, and a generation constraint can be tighter than a detection threshold.

Flags are humanizer's stable pattern IDs (see the table at the top of its SKILL.md), not display numbers, which can change between humanizer versions.

The general rule: a figure that could be dropped into any piece about any topic is not a figure, it's filler with parallel structure.

Guardrails

Both checks are mandatory before returning revised text.

Claim check

Figures compress, and compression eats qualifiers — but diction work compresses too, and it is applied "everywhere, always" with no budget to make you pause over it. Free does not mean unchecked. "Financial modelling indicates that X" tightened to "The math works" is a Saxon-default move — three Latinate words replaced by three Saxon ones — that also quietly drops the attribution: the claim goes from something a model produced to something stated as flatly true. That is the same failure the claim check exists to catch, just arrived at through diction work instead of a figure. Run the claim check against every sentence your revision touches, whether the touch was a figure or diction alone. Compare the claims in your revision against the original:

1. Did a hedge become a promise? ("an opportunity to improve" → "we will lead")
2. Did a qualifier vanish? ("in most cases," "for enterprise accounts," "after year two")
3. Did a number, date, or scope get dropped for rhythm?
4. Did an attribution disappear? ("financial modelling indicates" collapsing into an unattributed claim is exactly this, even when no figure was involved)

If yes to any: restore it, even at the cost of the figure. Flat prose that states the claim correctly beats a memorable sentence that overstates it. If the original claim was vague and you cannot sharpen it without inventing support, say so and leave it — don't manufacture a concrete detail to give the figure something to land on.

Ear test

Read the last five words of each paragraph aloud. Two checks:

• Stress. Does the final syllable land on a stress? Latinate endings trail off; Saxon monosyllables land.
• Breath. Does the closing clause fit in one breath? If it needs a second, split it or cut it.

Farnsworth's method is auditory throughout. A figure that only works on the page isn't working.

Revision checklist

1. Budget set? Length and register both checked, number written down.
2. Endings. Does each paragraph finish on a stressed Saxon word?
3. Latinate pileups. Broken up with verbs, not synonyms?
4. Key claim. Stated once abstractly, once concretely — with no transition phrase between?
5. Contrast. Does every figure have something to push against?
6. Triggers. Can you name the trigger condition for each figure you applied?
7. Budget respected? Count the figures. Over budget means cut, not justify.
8. Forbidden constructions. None present?
9. Claim check. Hedges, qualifiers, numbers, attributions all intact?
10. Ear test. Read the closings aloud.

Output format

Diagnose only: A findings list. Each finding locates the flat passage, names the figure its trigger licenses, and states the budget. Order by impact — endings first. Offer to apply.

Strengthen: Revised text, then a change note in this shape:

Budget: [length] → [N] figures
Applied: [figure] at [location] — trigger: [which condition fired]
Free: [diction changes, summarized in one line]
Left alone: [what you could have done and chose not to, and why]
Claim check: [any hedge/qualifier/number preserved or restored, or "clean"]

Single line: 3–5 treatments, each labeled with its figure, then a recommendation with one sentence of reasoning.

Always report what you left alone. Restraint is the deliverable as much as the figures are.

Worked example

Before:

The implementation of our customer experience transformation initiative will require significant organizational commitment and cross-functional collaboration. This strategic priority represents an opportunity to enhance our competitive positioning and drive sustainable growth.

Claims present: (a) it requires commitment and cross-team work; (b) it represents an opportunity — hedged — to improve competitive position and growth.

Overcooked — do not produce:

This transformation will test us. It will demand commitment. It will require every team to work together. But if we commit, if we collaborate, if we follow through—we will not just compete. We will lead.

Diagnosis: two anaphora runs in forty words against a budget of one. A negative parallelism ("not just compete"). A generic positive closer ("we will lead"). And the hedge, "an opportunity to enhance," became an unqualified promise. Six figures in one paragraph, four of them on humanizer's flag list, and a claim the original never made. This is what blowing the budget looks like.

Strengthened:

Customer experience programs fail for a boring reason: they need teams that don't report to each other to change how they work at the same time. That's the cost. Get it right and it could put us ahead of the competitors still selling on price. Get it wrong and we've spent a year on work nobody uses.

Budget: 68 words → 1 figure
Applied: antithesis at the close ("Get it right and... / Get it wrong and...") —
  trigger: both poles real, survives the swap test
Free: nominalizations to verbs ("the implementation of... will require" → "they need");
  Saxon finishes on "the cost," "on price," "nobody uses"
Left alone: no anaphora — under 300 words. No antimetabole — trigger tests don't fire.
  Isocolon considered for the three requirements and cut; they aren't equal in weight.
Claim check: hedge preserved ("could put us ahead," not "will"). "Cross-functional
  collaboration" made concrete as "teams that don't report to each other" — same claim.
  "Drive sustainable growth" dropped: the source gives no mechanism, and inventing one
  to feed the figure would be claim inflation. Flag for the author rather than restore.
  "Work nobody uses" adds a failure case the original only implied — confirm with the
  author before publishing.

Note the last two lines. The claim check runs in both directions: it stops you from inflating, and it makes you declare what you added.

Common issues

Zero figures is a real answer. If the text is clear, correctly hedged, and ends on a stress, you're done. Say that. Adding a figure to already-working prose is the most common way this skill makes writing worse.

The author's voice outranks the figure. If they write in short flat declaratives, a Churchillian triad is not an improvement, it's a costume. Match what's there.

Run humanizer first on AI-generated text. Figures layered onto promotional prose compound the problem. Clean the surface, then add craft.

Don't hunt for triggers. The trigger conditions exist to stop you from applying figures, not to help you find excuses. If you're reasoning hard about whether a trigger fired, it didn't.

Antimetabole is over-attempted. Most pieces should have none. If the three trigger tests don't fire cleanly, the output reads as a fortune cookie.

Some flatness is correct. A recall notice, a status update, a price change — these should be flat. Check the register table before doing anything.

Source

Ward Farnsworth, Classical English Rhetoric (2011) and Classical English Style (2020), plus his interview on David Perell's "How I Write" podcast.

Quotations are as delivered, condensed where noted. One exception: the active-voice version of Churchill's "so much owed by so many to so few" in references/figures.md is Farnsworth's own hypothetical foil, not a line Churchill wrote — it is labeled there.