Bullets
The smallest unit — and the one you use most
Craft unit
You write these constantly and had nothing written down about them. Every card you cut from a page, every list post, every feature row, every teaser line in an email is a bullet. A bullet is an unfinished thing with a benefit attached — not a feature, not a summary, not a shortened sentence. This page is how to build one on purpose.
Three sources converge here, which is why it earned a page. Schwartz's frame that copy is assembled from blocks rather than written; the unfinished magnetizer from the Hooks page, which explains why a withheld thing pulls; and your own card format, which turns out to be a bullet at display size. Feature-versus-benefit is universal craft. Nothing is reproduced from any source.
01
The unit — copy is assembled, not written
The useful reframe: a piece of copy isn't prose with lists in it, it's a set of blocks set into a structure. Bullets are the load-bearing block. Think of it the way the design system treats components — build the piece once, then move, cut, reuse and test the parts independently. It's why a good bullet survives being lifted out of the page it was born in.
hook
context
bulletbulletbulletbullet
proof
proof
the ask
The pink blocks are the ones that travel. Each can be cut out whole and become a card, a post, an email subject, or a feature row somewhere else.
The unit
Blocks, not sentences
Assemble the piece from parts you can move independently.
A bullet that only works in its original paragraph isn't finished. Portability is the test.
Yours: same mental model as the design system — components, not one long painting. The content pipeline already depends on this: cards get cut from pages.
02
The anatomy — benefit + loop
Two parts, and which one is missing tells you what's wrong. A benefit with no loop is a perfectly good sentence that gives the reader no reason to continue. A loop with no benefit is a tease about nothing. Note the rhyme with the hook anatomy — call-out plus value condition, benefit plus loop. Same shape, one scale down.
Benefit onlyYour agent will remember things between sessions.True, useful, and finished. Nothing pulls.
Loop onlyThe one thing nobody tells you about agent memory.Open, but promises nothing in particular. A tease about nothing.
BothThe storage choice that decides whether your agent remembers anything next week — and why most teams make it last.Names the payoff and withholds the mechanism.
The anatomy
Two parts, always
Something they want, in a shape that isn't finished.
Diagnose by subtraction: no pull means no loop, no interest means no benefit.
The loop is what's withheld — a name, a method, a number, a reason.
Yours: this is magnetizer 02 from the Hooks page at one-line scale. Same mechanism, smaller unit.
03
The ladder — same fact, three rungs
One underlying fact, climbed. This is the whole craft in three lines: rung one is what engineering hands you, rung two is where most copy stops, rung three is a bullet. The step from two to three is smaller than it looks — keep the benefit, take away the resolution.
FeatureStores conversation context in a graph database.What it is. Comparison fodder. Nobody buys here.
BenefitYour agent remembers what happened last week without you re-explaining it.What it does for them. Better — and complete, so they can stop reading.
BulletWhy the agents that remember last week almost never store context the way you'd expect — and the one-line change that fixes recall.Benefit intact, mechanism withheld. Now they have to keep going.
The ladder
Feature → benefit → bullet
Three operations on one fact, not three lengths of one sentence.
Feature answers "what is it". Benefit answers "what do I get". A bullet does the second while refusing to finish.
Write rung two first, always. You can't withhold a mechanism you haven't identified.
Careful: rung three is the rung nearest to clickbait. What keeps it honest is that the payoff is real and arrives — see job 07.
04
The loop kit — six things to withhold
Six named moves, so you're choosing rather than groping. Each withholds something different, and that variety is the point — a set of bullets that all withhold the same thing reads like one bullet written six times.
The withheld nameThe single setting behind most recall failures — and it isn't in the docs.
The method gapHow nineteen teams cut re-priming time by four fifths without adding a single service.
The counterintuitiveWhy adding more context made recall measurably worse.
The exceptionThe two situations where a graph is the wrong answer, even at scale.
The mistakeThe thing almost everyone does at step three that quietly undoes step two.
The thresholdThe point where flat files stop working. It arrives earlier than anyone plans for.
The loop kit
Six ways to leave it open
Name, method, reason, exception, mistake, threshold.
Vary the type across a set. Same withholding six times reads as one bullet repeated.
The exception and the mistake are the most honest of the six — both promise something genuinely useful rather than merely absent.
05
The rules — few, parallel, varied, specific
Four disciplines. The first is the one people resist and the one that matters most: bullets devalue each other, so more is not better. The rest are about making a set read as a set.
Few
Four excellent bullets beat fifty adequate ones. Every extra bullet is a claim on attention that the next one now has to earn back. A long list reads as filler even when every line is true.
Parallel
Within a set, keep the grammatical shape consistent. The eye reads a set as one object, and a broken shape reads as a mistake rather than as variety.
All starting with "The…" or all with "Why…" or all with "How…" — pick one per set.
Varied in length
Same shape, different lengths. Six bullets of identical length drone; the rhythm is what carries a reader past the third item. Long, short, short, long.
Specific
Concrete nouns and non-round numbers. "Four fifths" and "nineteen teams" land where "dramatically" and "many" evaporate — and vague bullets read as invented, because usually they are.
The rules
Four disciplines
Few, parallel, varied in length, specific.
Cut to the four best. The cut ones aren't wasted — they're candidates for the next piece.
Read a set aloud. Parallelism failures and length droning are both audible before they're visible.
Yours: specificity is the proof ladder arriving at one-line scale. Same rule, same reason.
06
Your card is a bullet at display size
The pipeline already does this without naming it. A card cut from a page is a bullet rendered large — benefit plus loop, one line, standing alone. Which means card-making is a copy task wearing a design costume, and everything above applies to it directly.
Signal Thinking
Adding more context made recall worse.
— the counterintuitive
Signal Thinking
Flat files stop working earlier than anyone plans for.
— the threshold
Both are loop-kit types set at display size. The loop closes in the linked piece rather than on the card — which is fine, as long as the link day actually arrives.
The card
Same unit, bigger type
Cut cards using the loop kit, not by picking a nice sentence.
A card's loop closes one click away, not on the card. That's legitimate — but it makes the link day a promise, not an option.
Three card days then a link day means three open loops outstanding. Track which ones you've closed.
Yours: reframes card-cutting from a design chore into the copy move it already was. Full pipeline detail lives in
the Content Engine.
07
Failure modes
Four ways this goes wrong. The first one isn't a craft problem, it's an honesty problem, and it's the reason this page has a hard rule rather than a preference.
The loop that never closes
You opened it and the piece ends without resolving it, or resolves it with something trivial. That's clickbait regardless of intent. It works exactly once per reader, and it spends the trust every future hook depends on.
FixBefore publishing, list every loop you opened and point at where each one closes. Any that don't, either close them or cut them. This is the hard rule of the page.
All feature, no benefit
The list reads as a spec sheet. Common when bullets are written from a changelog, because a changelog is rung one of the ladder and nothing pulls it up.
FixRun each line through the ladder. If you can't say what it lets the reader do, it isn't a bullet and probably shouldn't be on the page.
All loop, no substance
Every line teases and none of them promises anything specific. Reads as evasive, because it is — usually a sign the underlying thing isn't differentiated enough to name.
FixWrite rung two for every line first. If the benefit is thin, that's a positioning problem showing up as a copy problem. Go upstream.
The wall of fifty
Quantity used as a substitute for quality. Each bullet devalues the ones around it, and past about six the reader stops reading them individually and starts skimming the shape.
FixCut to four. Keep the rest in the swipe file — they're the raw material for the next piece, not waste.
Failure modes
One honesty problem, three craft problems
Only the first is a refusal. The rest are fixable in an edit.
Hard rule: every loop you open closes inside the piece, or one click away with the click delivered. No exceptions — this is the line between a bullet and clickbait.
Two of these are symptoms of upstream problems. Thin benefits mean thin positioning; a wall of fifty usually means no one decided the one thing.