Collect the writing that was actually for someone
Step 125 minutes
By the end of this step
When you finish this step you will have one plain text file of your own prose, one reader, one register, with a word count written on it.
Words used in this step Corpus · Register · Plain text · wc -w · Model
- Corpus
- The one file of your own writing that everything later is measured against. Chosen by hand, piece by piece, never exported in bulk.
- Register
- The way you write for a particular reader at a particular speed. A two-line reply from your phone and an edited email are two registers.
- Plain text
- Words with nothing else attached. No fonts, no bold, no pictures. The file name ends in .txt.
- wc -w
- A command that counts the words in a file. You type it into a terminal, the window where a computer takes typed instructions.
- Model
- The AI doing the writing, the engine behind a chatbot. It copies the patterns in whatever you show it, good or bad.
Sixty thousand words go into the folder in one lump. Somebody decides to teach a model how they write, so they export everything they have ever typed. Four years of chat messages, a wedding speech, the invoice notes, a cover letter, the two posts that did well. What comes back is grammatically theirs and belongs to nobody.
That is not the model failing. It is an average, and it did exactly what it was asked. The average of a person writing fast and the same person writing carefully is a third person who has never existed.
The instinct underneath is that more input makes a sharper picture, and it does the exact reverse. Every extra register drags the numbers toward the middle, and the middle is the flat competent register you were trying to get away from.
Export everything
- Four years of chat messages
- A wedding speech, a cover letter, invoice notes
- Sixty thousand words in one lump
- Comes back grammatically yours and nobody's
One reader, one register
- Pick the reader the output has to land with
- Eight to fifteen pieces that actually went out
- Everything you did not write cut
- 1,500 to 3,000 words, counted
The rules
You are not collecting your writing. You are collecting the writing you did for one reader. Pick the audience the output has to land with. Customers, peers, the people on your list, strangers who arrived cold. Everything you wrote for them goes in. Everything you wrote for anybody else stays out, however good it was.
Your fast register and your deliberate register are two different writers. The reply you fired off in ninety seconds and the piece you edited four times do not share a rhythm. Both are you, and only one of them is the one you want reproduced at scale, so you have to decide which before you collect anything.
Published beats drafted. Writing that actually went out got the pass where you cut the clever bit and fixed the opening. A draft you abandoned is a record of you failing at your own standard. Teaching a model from your failures is a peculiar plan.
Two thousand words is enough. Twenty thousand is worse. You are measuring shape, not accumulating evidence. Eight to fifteen real pieces will do it. Past that you are adding noise and quietly widening the range the model is allowed to write inside.
Cut everything you did not write. Quoted paragraphs, pasted client copy, the boilerplate at the bottom, headings a template handed you, anything a colleague rewrote. It is your file and it should contain nothing but you. Do this before you count a single word.
Do it now
Make a folder. Inside it, one file called corpus.txt, plain text, no formatting furniture.
Open the place your real writing lives and pick eight to fifteen pieces that went to the reader you chose. Paste them in, separated by a blank line. Strip the quotes, the signatures and anything somebody else touched.
Then put one sentence at the top of the file, above everything. Name the reader and the register in that sentence, not the topic. Something closer to this shape:
Written for people who already run a business and are deciding whether to trust me. Deliberate register, edited.
A worked example
Say you are a solo wedding photographer. You write every client email, every enquiry reply and every blog post yourself. The job you want a model drafting is the enquiry reply, so the reader is a couple who has enquired and not booked yet.
Everything written for anybody else stays out. That includes the blog, however well it did.
Worked example
A solo wedding photographer
Your corpus for enquiry replies, built from what actually went to couples who had not booked yet.
First line: Written for couples who have enquired and not yet booked. Deliberate register, edited. In: 9 enquiry replies that went out, each written fresh In: 3 emails sent after a first video call with a couple Out: blog posts, written for strangers arriving from a search Out: gallery delivery emails, written to couples already booked Out: two-line replies sent from the phone between shoots Cut: signatures, the couple's own questions quoted back, the pasted price list wc -w corpus.txt 2140 corpus.txt
What to notice
Your best piece is a blog post about a wedding in the snow. It went to a different reader, so it sits in a second file for later. It is not in this one.
Twelve pieces, one reader, 2,140 words. Inside the range.
The failure you will hit
You will want to include the best thing you ever wrote even though it went to a different reader. It is the piece you would show somebody. Leave it out. A corpus is not a portfolio, and one brilliant off-register piece in eight will move your numbers further than four ordinary on-register ones will hold them still. Put it in a second file and build a second profile later if you want it.
Not for you if
You write for one person who already knows you, and the rhythm never has to survive a stranger reading it cold. Building a measured profile for an audience of one is work you will never earn back.
That is the corpus built. Step 2 turns it into numbers. Content & Copywriting is one of the Vault's categories.
Done means
wc -w corpus.txt returns a number between 1,500 and 3,000, and reading the first line aloud names a reader rather than a subject. Run it again after every deletion. If the count falls under 1,500 you have collected too narrowly, and if it climbs over 3,000 you have stopped choosing.
I keep this tick in this browser and send it nowhere. It goes when you close the browser.
Untick it if the check stops passing.
This step files under Content & Copywriting in the Vault.
Open Content & Copywriting
Sports Tech Nation
Included in Membership
The next step is included in Membership.
This step is open to everyone. The rest of the track is not.
Step 2, Measure it, do not describe it
What this step installs
When you finish this step you will have nine numbers describing your own prose, computed from your corpus file, saved somewhere you can paste them into a prompt.
20 minutes
It hands off to Content & Copywriting in the Vault, which the same membership opens.The track list stays open to anyone.
Included in Membership — £79/month.
You sign in first, so the membership lands on the right account.
Your ticks in this browser stay. They come with you when you join.