Hey, Mark Masters here.
I asked Claude for twenty bullets last Thursday afternoon, got twenty back, and read them out loud at my desk.
Every one was a clean sentence. Every one said something true about the product. Not one of them made me want anything.
It took me an hour to work out what was wrong, and the reason it took an hour is that nothing was wrong with them as sentences. They were good. They were just the wrong unit.
Here is the build I put together afterwards, and the count from running it twice.
The Gap Is the Product
A bullet is the only thing you write in direct response whose job is to not tell you.
It opens a question the reader cannot answer, and the only place that answer lives is inside the thing you are selling. Delete the answer from a sentence and you have made a bullet out of it.
That is the opposite of what a model is built to do, which is most of why this goes wrong.
I ran the whole thing on a sample offer I put together for the test. A $497 course for wedding photographers who book solid from May to October and then watch the calendar go quiet until March.
Eight mechanisms inside it, each one with a real answer behind it, because a gap with nothing behind it is a different kind of failure and we will get to that.
Three sessions, run separately:
One wrote twenty bullets from an ordinary request.
One wrote twenty under the system below.
One graded both sets, having seen neither brief and neither offer.
Eleven of the first twenty were dead on arrival. Here is the test that killed them.
A Bullet Fails When the Reader Can Answer It
Take any line in your bullet block. Ask whether the reader can satisfy the curiosity it raises using only that line and what they already know.
If they can, it is a feature line.
The gap is not decoration on top of the bullet. The gap is the bullet, and everything else in the sentence is scaffolding holding it open.
Three pairs from the same course, same fact each time. The left column is what Claude handed me cold. The right column is the same fact with the answer removed.
The fact, told | The same fact, withheld |
|---|---|
Send the venue coordinator a one-page PDF of the ten weddings you have shot at their venue, because coordinators rarely have current photos of their own rooms. | The one-page PDF that puts you on a venue's preferred-supplier list, and the thing coordinators are quietly short of that makes them keep it. |
Ask the second shooters who worked beside you for referrals, because they have seen your work and hold overflow dates they cannot cover. | The referral source that has watched you work a full day, has dates they cannot cover, and has never been your client. |
Email the couples who enquired and never booked 11 months later, when their friends start planning. | Why the eleventh month after a dead enquiry is worth more than the first week. |
Read the left column and you are finished. You know the move. You might even go do it.
Read the right column and there is a hole in the middle of each sentence that only one thing fills.
Notice what the right column keeps. The specificity is all still there, and in a couple of places it is sharper, because "the eleventh month" arrives without its reason attached.
Caples was right that "$50,000" beats "a fortune," and the same instinct applies here. A specific gap is worth more than a vague one, because the reader can feel the exact size of what they are missing.
Why the Model Closes Every Gap You Open
A request for bullets looks, to a model, like a request for information about a product.
So it gives you information about the product. Completely. Helpfully. In whatever number you asked for.
I am reasoning about behaviour here rather than reporting anything I can prove about what happens inside the model. What I can describe is the pattern from the outside.
In almost every context a model works in, withholding the answer is a defect. It is the thing that gets a response marked as unhelpful.
Nova found the same shape from a different angle in August, grading copy with the Lever. The model produces what the request looks like it wants.
You ask for | What the request looks like | What comes back |
|---|---|---|
A score out of ten | A grading job | A number, and a rewrite that chases it |
Twenty hooks | A variations job | One idea, twenty times over |
Twenty bullets | A product-information job | Twenty complete, correct sentences |
The one place where withholding is the deliverable is this small corner of sales copy, and nothing in your request tells it that.
You are fighting a strength that is pointed at the wrong target.
Which is why the fix is not a better adjective in the prompt. It is a rule with a pass condition, and a second reader who applies it.
Write the Payoff First, Then Take It Out
Before any prompt, this is the move by hand, and it is the part of this issue that works with no model in the room.
Write the complete sentence. The one with the answer in it. The feature line.
Then take the answer out and keep the shape.
That is it. The sentence you cut is the sentence that proves you had something to withhold, which is why writing it first matters more than it looks.
The missing piece sits in one of four places.
Withhold the what, and name the result instead: "The one-page PDF that puts you on a venue's preferred-supplier list."
Withhold the how, and name the mechanism instead: "The referral loop that runs entirely through people who will never compete for the same booking."
Withhold the who, and name the move instead: "The people who watched you shoot all last season and have work they need to hand to someone."
Withhold the when, and name the action instead: "Why the eleventh month after a dead enquiry beats the first week."
That is not a taxonomy and I am not going to grow it into one. It is four places to look when a bullet is coming out flat, and if you have written the answer sentence first, the flat one usually fixes itself in about fifteen seconds.
Two of those four are where the gap should sit most of the time, in my experience. The what and the when. The how tends to produce bullets that sound like a process, and the who works best when the person is genuinely surprising, the way second shooters are surprising to a photographer who has only ever asked past clients.
Build the Writer, Then Build the Executioner
Now the system. Two prompts, two separate sessions, and the separation is the whole thing.
The first writes bullets under the withholding rule.
You are writing bullets for a sales page. A bullet's job is to open a gap the reader cannot close without the product. A sentence that answers its own question is a feature line, and it has failed.
Method for every bullet: write the complete sentence, the one that contains the answer. Then remove the answer and keep the shape. The missing piece can be the what, the how, the who or the when.
Rules:
- A bullet whose curiosity the reader can satisfy from the bullet itself has failed. Cut it and write another.
- Every gap must have a real answer in the material below. Do not open a gap the product cannot close.
- Do not name a mechanism and explain it in the same line.
- One sentence per bullet.
Write [NUMBER] bullets for this offer.
[Paste the offer: every mechanism, what it does, and the reason it works. The reasons matter most. They are the answers your gaps will be built on, and the model cannot withhold what it was never given.]The second prompt goes into a fresh session that has never seen the first one, never saw the offer, and does not know who wrote the lines.
Below are numbered lines from a draft sales page. A stranger wrote them. Judge each one on its own.
For each line, answer one question: after reading only this line, what does the reader still not know?
If the answer is "nothing" — the line tells the reader everything it raises — mark it DEAD.
If something specific is still withheld, mark it LIVE and name the missing piece in a few words.
Return one verdict per line in this format: `3. DEAD` or `3. LIVE — the missing piece`. End with the count of LIVE and DEAD.
[Paste the bullets, numbered, and nothing else. No offer, no brief, no context.]Two things about that second prompt are doing the work.
It asks about the reader rather than about quality. "Is this a good bullet" gets you an opinion. "What does the reader still not know" gets you an inventory, and an empty inventory is a verdict you cannot argue with.
It has no authorship to defend. A session that just wrote twenty bullets will grade them generously, which Nova demonstrated on a different artifact in August. A session that thinks a stranger wrote them has nothing invested.
That second one is why pasting the grading prompt into the same chat window gets you softer verdicts.
One more rule, and it is the one people break: send the bullets with no context at all. Paste the offer alongside them and the grader starts answering the questions out of the offer sheet, marking gaps closed that your reader could never close.
What the Gate Killed, Both Times I Ran It
Here is the count.
Cold, no rule | Written under the rule | |
|---|---|---|
Bullets in | 20 | 20 |
Marked LIVE | 9 | 20 |
Marked DEAD | 11 | 0 |
Eleven of twenty from an ordinary request, dead. The cold set was not badly written either.
Several of the dead ones were the most quotable lines in the batch, including a nice antithesis about eighty photographs proving you have shot a lot of weddings and eight proving you can shoot theirs.
That line is good. It is also finished, and there is nothing left to buy.
Here is what it looks like rebuilt. The fact underneath is that couples want to see one wedding covered properly rather than a highlight reel of ten.
Dead: "Eighty images prove you have shot a lot of weddings. Eight prove you can shoot theirs."
Live: "What a couple is actually looking for in your portfolio, and why most galleries are built as if they wanted something else."
The rebuilt version is less satisfying to write. That is the trade, and it is worth naming, because the lines you are proudest of are the ones most likely to die on the gate. A sentence that closes cleanly feels finished to the writer for the same reason it feels finished to the reader.
The nine cold survivors are worth a second look too, because most of them survived by accident. They were bullets that happened to run out of room before they got to the answer, rather than lines built to hold something back.
The second run needs a caveat I would rather print than bury, because twenty for twenty is a suspiciously clean number.
The grader flagged three of those twenty as thin without being asked to.
One gave away the tactic and kept only the reasoning.
One named nothing specific at all.
One spent two-thirds of its length listing modules before it opened anything.
So the gate sets a floor. It catches the bullet that answers itself, and it passes the bullet that withholds something nobody wanted in the first place. The standard above the floor is still yours.
Three more things the gate does that you should expect.
It kills good feature lines. A feature line is a different unit with a different job, and a spec block full of them is correct. Run the gate on your bullet block only, then put the survivors of the cut back in the spec where they belong.
It passes manufactured mystery. A gap with nothing behind it reads as LIVE to the grader and reaches the reader as a lie, because they click through and find no answer.
Before you write a bullet, you should be able to say its answer out loud in one sentence. If you cannot, what you have is a tease, and the person who buys on it is the person who asks for a refund.
It encourages volume. Thirty bullets is a page with no argument, and bullets are the easiest unit to overproduce once a model is generating them. Generate forty, gate them, and ship the eight that carry the eight things your offer actually does.
Turn the Gate Into a Skill You Never Have to Paste
The gate is one rule with a clean pass condition, which makes it a better Claude Skill than a prompt.
A prompt works when you remember to paste it. A skill fires whenever you are writing bullets, in any project, without you doing anything.
That is the argument for packaged assets in general, made small enough that you can build one yourself this afternoon.
The spec is four parts.
The description that fires it: "Use when writing bullets, fascinations, or benefit lines for a sales page, email or VSL." That line is what decides whether the skill triggers, so write it in the words you would actually use.
The pass condition: a bullet passes when a reader cannot satisfy the curiosity it raises from the bullet alone, and fails when they can.
The grading pass: one question per bullet, which is what the reader still does not know. Nothing means DEAD.
The reporting rule: report the kill list rather than silently rewriting, because the dead ones are usually telling you something about the offer.
Build it, use it for a month, and you will have a far better sense of what a packaged skill is worth than any argument I could make about it here.
The Read-Aloud Test, Which Needs No Tools at All
Read the bullet out loud and stop at the end of it.
If you feel finished, it failed.
That is the whole test, and it is the one I use on my own lines before anything goes near a second session.
The gate catches what you miss. Your ear should catch most of it first.
The Master’s Memo
The model is good at being complete, and complete is fatal in this one place.
Write the sentence with the answer in it. Then take the answer out, and make sure you can still say it out loud when a buyer asks.
More clicks, cash, and clients,
— Mark Masters
