Did you write this for them, or did you bring a written?
In breaking, one of the things you're judged on is whether you danced to the song that was playing. Not whether the moves were good. Whether they were for this song, right now. Bring a rehearsed routine and run it regardless of what the DJ does, and the judges see it immediately no matter how clean the execution is.
Most things people send that claim to be personal are rehearsed routines. The recipient can always tell, and it costs more than sending nothing, because it asks for credit for effort that wasn't spent.
freestyle runs the test on your cover letter, cold email, proposal, pitch, or toast before it goes out.
Change one variable, see what breaks. Five axes: the name, the organization, the domain, the occasion, the sender.
Any sentence that survives all five was never about them. The skill marks every survivor, quotes them back as one block, and gives you the ratio. Reading your own template assembled is more convincing than being told about it.
The axis nobody passes is the last one. If your message works identically coming from any competent stranger who did the same research, you put nothing of yourself in it.
Not all specificity counts. There's a ladder, and most work stops on the bottom rung believing it arrived.
Tier 1, merge field. Their name, their company, their funding round. Anything a script could insert. This scores as generic here, on purpose, and that's the whole argument. Your recipient has seen hundreds of these and the pattern is unmistakable. Visible fake attention performs worse than no personalization at all, because a plainly generic short message doesn't insult anyone.
Tier 2, research. You went and read the thing. Real effort, and any competitor willing to spend the same hour could write your sentence.
Tier 3, observation. Something only someone paying attention would catch. A tension between two things they said. The thing everyone praises that's actually the least interesting part. This is the first tier that gives the recipient something instead of asking for credit.
Tier 4, stake. Something only you could say. A position, a disagreement you're willing to have, what it costs you if you're wrong. It can't be researched, which is why it can't be faked.
Then the second axis, which almost nobody checks: could this have been sent six months ago? A message can be deeply specific to a person and still dead, because it responds to who they are instead of what's happening.
The written. It rewrites your message as the mail merge it actually is. Your surviving sentences verbatim, merge fields exposed where the real values sat, same structure, no commentary.
Hi [FIRST_NAME],
I came across [COMPANY] and was really impressed by [RECENT_ANNOUNCEMENT].
...
11 of 14 sentences survived. This is what your recipient reads.
Then it stops. It doesn't pile on. You're already looking at it.
git clone https://github.com/MichaelDoesMarketing/freestyle.git ~/.claude/skills/freestyleOr grab the packaged .skill from releases.
"Cover letter for a job I actually want. Be honest with me."
"Nobody replies to this email. Here it is."
"Best man speech. Does it sound like I know him?"
Four modes. Swap test for the full read. The written when you can take it. Off the dome when there's no draft yet and you want the material before you write. Range check for anything going to a list, where the honest answer is what to make specific and what to accept as shared.
I wrote release notes for six repos in this portfolio earlier today. Real text, published, still live. So that's what it ran on.
THE SWAP
Every one of them is three sentences. Here's what survives all five swaps, quoted from the live releases:
"First release." Survives everything. Identical across all six. Pure template.
"Install: download [name].skill and open it, or clone the repo into ~/.claude/skills/" Technically breaks the name swap. That's a false break, and it's the exact trap the skill is built to catch. The filename is a merge field. The sentence around it is unchanged in all six.
What breaks honestly is the middle sentence, and only the middle sentence. "Every item gets a job, every adjacent pair gets the swap test" cannot be moved to another repo. That one's real.
Ratio: 2 of 3 sentences are template. 67%.
THE CEILING
Tier 2 across the board. The middle sentence is accurate and researched. Nothing reaches tier 3, and nothing has any stake in it. A release note written by a stranger who read the README would be identical.
THE VERDICT
Mostly correct. Install instructions should be shared, and pretending otherwise would be worse. But "First release." is dead weight opening the most-read line of the page, and it's doing nothing that the version number doesn't already do.
The fix is deleting two words from six pages, which is what a good pass usually produces. Not more writing. Less.
Worth saying: this ran on my own output, in someone else's repos, and the honest ratio was 67% template. A skill that only finds problems in other people's work isn't a skill, it's a weapon.
It will never invent specificity. If nothing real is there, it says so. "I've followed your work for years" from someone who hasn't is the fastest way to lose a reader who checks.
It won't demand the impossible. Some things go to a thousand people. The useful answer is what to make specific and what to let be shared, and fewer recipients at tier 3 beats everyone at tier 1.
Not everything should be a freestyle. Confirmation emails, invoices, legal notices, safety information. Those should be identical for everyone, and it says so instead of forcing the frame.
And it won't run the written on a condolence note or an apology. That mode is for outreach and applications, where the stakes are professional and the discomfort is useful.
battle-tested writes the line. versus picks the fight. subtext reads what came back. sequencing puts it in order. sampled checks whose it was. headroom leaves room for whoever has to use it. freestyle asks whether you made it for this room at all.
MIT. Use it, fork it, ship it.