Hey, Mark Masters here.

Most copywriters use Claude or ChatGPT well at the drafting stage and badly at the editing stage.

They write a draft. They paste it in. They type "make this stronger" or "edit for clarity and flow" and hit return. They get back a version that's smoother and stripped of about thirty percent of whatever made the draft theirs in the first place.

That's laundering.

I built a five-pass editing system to fix it. Each pass does one job. Each one runs in a specific order. The output of each one feeds the next. Used together, the Protocol does what a single "edit this" prompt cannot do, which is catch the five different problems your draft probably has without flattening the parts you got right.

Here's the full build.

There are two failure modes most working copywriters fall into when they try to use AI at the editing stage.

The first is the vague single pass. "Make this clearer." "Tighten this up." "Edit for flow and persuasion." Each of those is asking the model to do five different jobs at the same time, which means it does none of them well. It produces a version of your copy that reads more smoothly because it's been homogenized, and any specific edits worth keeping are buried under three rounds of unnecessary rewrites.

The second is no AI at the editing phase at all. You used Claude to draft. You edit by hand. Whatever speed gain you got at the drafting stage gets eaten back at the part of the process where you most need a second pair of eyes that doesn't share your blind spots.

Both failure modes have the same root cause. Editing is at least five jobs. Asking a model to do all five in one prompt produces output that does none of them properly.

The Protocol replaces the vague single pass with five lenses, each defined to look for one specific kind of problem and report back. You run them in order. You read each pass. You apply the edits you agree with and skip the ones you don't. The whole thing takes about twenty minutes on a 1,500-word draft.

Each lens is a prompt with a narrow job. Specifying the job is what makes the pass useful, because the model now has criteria to judge against instead of guessing what "stronger" means.

The order matters because each lens depends on the one before it.

You can skip a lens if it doesn't apply. A short ad probably doesn't need objection handling as a separate pass. A draft with no brand voice samples available has no way to run the voice match lens cleanly. The Protocol is a sequence, not a checklist. Use the lenses you need.

What this pass is looking for: sentences that ask the reader to do too much work. Stacked ideas. Pronouns with unclear antecedents. Industry shorthand that snuck in. Any sentence where you'd have to read it twice to know what it's saying.

A few things about why it's structured this way.

The "clarity only" boundary is doing real work. Without it, Claude drifts. It will start "fixing" persuasion or voice while it's supposed to be looking at clarity, and you'll lose the focused signal that makes this pass useful. The constraint comes from naming what the pass should ignore.

The "do not manufacture issues to hit a number" instruction matters more than it sounds. Claude has a strong bias toward giving you a satisfying-feeling list of edits even when the draft is mostly clean. The first time I ran this pass without that line, I got five "issues" on a draft that genuinely had two. The other three were stylistic preferences dressed up as clarity problems.

The "written in the same voice as the rest of the draft" instruction prevents the most common failure of this lens, which is suggested fixes that read like a different writer wrote them.

This pass looks for the most common structural failure in B2B and B2C copy alike. Sections that explain a feature when they should be selling a benefit. Sections that prove a claim that nobody disputes while leaving the disputed claim unsupported. Sections that build credibility for a third paragraph instead of asking for the click.

The diagnostic-only framing is the part that took me a few attempts to land. Earlier versions of this prompt asked for suggested rewrites alongside the diagnosis. The rewrites were almost always weaker than the original because Claude was guessing at the strategic context.

When I removed the rewrites and made the lens purely diagnostic, the output got better. I read what's flagged, agree or disagree, fix what I want to fix, and move on.

The "performing rigor instead of moving the reader toward the action" line is worth tuning to your project. For long-form sales copy, that line is exactly right. For technical B2B copy where rigor is what builds credibility, you'd reverse it. Adapt it to the brief.

This lens only runs cleanly if you have voice samples for whoever the copy is supposed to sound like. The brand. The founder. The persona you've defined. Without samples, this pass is asking Claude to compare the draft to a description of a voice rather than the voice itself, and the output is unreliable.

If you have samples, the pass looks like this.

The reason this lens needs samples and not just a description is that Claude's interpretation of "founder voice, casual but authoritative" is going to be a flattened average of every founder who ever wrote casually. The samples force comparison against the actual voice instead of the model's idea of it.

How many samples? I've found three to five works for most projects. More than five and Claude starts averaging across them in unhelpful ways. Fewer than three and one outlier sample skews the comparison.

If you don't have samples for this client yet, skip this lens and add it back next time. Don't fake it with a description.

This pass looks for objections the draft hasn't addressed and objections it addresses badly. It's the lens that most consistently catches things I missed.

The "default to skepticism" line matters. Without it, Claude tends to validate the draft. It reads what's there, agrees with the framing, and finds the objections the draft already raised, which isn't useful. With the skepticism instruction, it acts more like a hostile reader, which is the lens you actually want.

The audience description has to be specific. "Marketers" is too broad. "B2B SaaS marketing managers at companies between $5M and $20M ARR who have tried two or three AI copywriting tools and been disappointed" is the level of specificity that gets you a useful objection list.

The last pass. Cut what isn't working. Sharpen what is. Replace weak verbs. Kill modifiers that aren't earning their keep.

The "10-15% word count reduction" target is calibrated. Without a target, Claude either does nothing or chops the draft in half. With a target, it makes proportional cuts.

The "do not cut anything that does load-bearing work" rule is the one I added after watching this pass remove the connective tissue from a long-form sales letter. The tightening lens has a bias toward density. Sometimes copy needs space to breathe between heavy sections, and the load-bearing work that space is doing isn't visible if you're scanning for cuts.

Read this output more carefully than the others. Maybe sixty percent of the suggested cuts will be good. The rest will be removing the parts that make the copy yours.

One lens at a time. Not all five at once.

Open a fresh chat for each lens. Paste the draft. Run the prompt. Read the output. Apply what you agree with directly to the draft. Move to the next lens with the updated draft.

The reason for fresh chats is contamination. If you run all five lenses in one conversation, the model carries assumptions from earlier passes into later ones. The clarity pass concluded the draft was clean, so the persuasion pass weights more lightly. The voice match pass already rewrote a section, so the objection pass evaluates the rewrite instead of the original. Fresh chats keep each lens uncontaminated.

What if two lenses give conflicting suggestions? They will. The persuasion lens will tell you a section is weak and needs more proof. The tightening lens will tell you the same section has too much proof and should be cut. That's the system working as designed. Two different jobs are flagging the same passage for different reasons. Your judgment decides which one wins. The Protocol surfaces the tension, not the answer.

Three failure modes I've hit using this on real client work.

The voice match lens flattens idiosyncrasies. If your client has a tic the AI reads as a "drift" — say, a one-sentence paragraph that breaks the rhythm on purpose — the voice match lens will flag it.

Sometimes that flag is right. Sometimes the tic is the thing that makes the voice work.

You have to know the voice well enough to overrule the lens. If you don't know the voice that well yet, this lens is not yet ready for that client.

The objection handling lens can overreach. If you give it an audience description that's too broad, it'll generate objections that the actual audience would never raise. "Some readers may worry about long-term commitment" is a generic objection that sounds plausible until you check whether your audience has ever cared about long-term commitment in this category.

They probably haven't. Tighten the audience description until the objection list stops including filler.

The tightening lens cuts your voice. I covered this above but it bears repeating because it's the most common failure of the whole Protocol. The last pass is the one most likely to strip what makes the draft yours. Read it last and read it slowly. Reject more suggestions than you accept.

A 1,500-word draft will take about twenty minutes to run through all five lenses. A 4,000-word sales letter will take closer to an hour, mostly because reading the output of the persuasion and objection lenses takes longer when the draft is bigger.

You will accept maybe sixty to seventy percent of the suggestions across all five lenses. The rest you'll skip. That's correct. A protocol that produced a hundred percent acceptance rate would mean Claude is rewriting your judgment instead of supporting it.

The drafts you ship after running the Protocol will be tighter, clearer, and more honestly persuasive than the ones you'd ship after a single "make this stronger" pass. They will not be drastically different, which is the point.

The Protocol sharpens drafts you wrote. It is not a system for getting Claude to write the draft for you.

Build it once. Save the five prompts somewhere you can grab them in thirty seconds. Run them on the next draft you'd otherwise have edited blind.

One pass cannot do five jobs. Five passes, each defined narrowly, can.

The Protocol takes twenty minutes on a typical draft and catches the kinds of problems you've gone blind to by the time you finished writing.

Run it on the next thing you ship and tell me what it caught.

More clicks, cash, and clients,
Mark Masters