April 22, a Wednesday morning. Dirty chai with coconut fat on the desk, a leftover from fading keto diet experiment. Obsidian vault open on one screen, four of my own BIP posts on the other, and a one-paragraph prompt I'd been working on for a week. The prompt told Claude to be a Paris Review editor reading my work against David Sedaris at his best. I lifted the "Paris Review editor" framing from a model-prompt example so I wasn't writing a prompt that would tilt back into sycophancy. Sycophancy is the thing I'm trying to cull as thoroughly as possible. (Wrote about that two days ago.) Don't tell me what's good. Tell me what a ruthless editor would cut. Score me 1 to 10 on hook, scene, specificity, rhythm, ending.
I hit return on Day 12 first because I knew it would be the worst of the four. Day 12 was the post I wrote on the third anniversary of Finn's death. Hallmark-saturated. Beautiful angel. Guiding love and light. Lives changed permanently. Every safe abstract noun I'd reached for instead of writing one specific thing he had actually done.
The score came back at 2.8 out of 10.
Yikes. 2.8 on a post about my son's death. The raw emotion of writing about something that traumatic, and the result is text-book Hallmark moment. I sat with that for a minute. Why would a Paris Review editor flag a deeply personal topic that hard?
Then I read the line in the critique that said: the cruelty that killed Finn wasn't cosmic. It was a train, a dark station, budget constraints at SBB, the main Swiss train provider. Get specific or cut.
Brutal, and all true. I had heard rumours that the lights at that station were cut at 11pm to save energy, the station being too small to justify the running cost. The cruelty was a budget line in a transport authority's spreadsheet. Putting the universe took him on the page hides all of that.
I ran the prompt against the other three posts. Scores below, on hook, scene, specificity, rhythm, ending, then the average.
- Day 12, Una Voz Delegada (Finn's third angelversary). 3, 2, 3, 4, 2. Avg 2.8.
- Day 18, Fat Skills, Thin Harness. 4, 4, 5, 4, 6. Avg 4.6.
- Day 26, Write Like Nobody's Reading. 6, 4, 5, 5, 5. Avg 5.0.
- Day 23, Patrouille des Glaciers. 5, 7, 7, 6, 6. Avg 6.2.
Humbling, but exactly what I asked for. One comment lodged in particular: my writing sounded like an MFA workshop. (For people not in the writing world: an MFA workshop is the standard graduate-school setting where ten people sit around a table polishing each other's prose into the same competent neutral register. Polished, plausible, a little dead.) Fine. I can handle that. The reason I'm doing this whole BIP thing is to get better at writing. I've always called myself a 5 out of 10 writer. Maybe in a year, with a lot of practice and this kind of feedback, the grade moves to a 6.5, maybe even a 7. Plenty of authors I admire are at that grade and augment their writing with the material, the research, and the ideas they spend years developing. That is what this BIP is really about.
I copied the whole critique into a file in the vault, named it BIP-Sedaris-Critique-2026-04-22.md, and started building the skill that afternoon.
Why Sedaris
Why him and not someone else.
He renders silly action flat. He never reaches for a pop-culture simile to announce that he is being silly. The action is silly. The sentence reports it.
He is the opposite of AI-fine. AI-fine defaults to general statement plus stacked adjective. Sedaris defaults to one named object.
And then there is Now We Are Five, his New Yorker piece from 2013 about his sister Tiffany's suicide, collected later in Calypso. Sedaris doesn't write Tiffany's death. He writes the family beach trip after. The piece ends on his other sister Gretchen asleep on the couch, with Rick Moody's The Ice Storm face-down on her chest. The grief is in that book on her chest. There is no beautiful soul. There is no she will be missed. There is a paperback on a sleeping woman.
That image is the proof. You can render the worst thing that has ever happened to your family without a single Hallmark word. Every time I write about Finn, I immediately start thinking of the first things that come to my mind, which are things like: beautiful soul, shattered hearts, the darkest stretch, the universe took him. None of those is true. The cruelty that killed him was a train, a dark station, and a budget line. Sedaris is the bar that keeps me on the specific.
Reference books for the ear: Calypso and When You Are Engulfed in Flames.
The prompt, in case you want to lift it but choose your own author and writing style to grade it
> You are a Paris Review editor. Read this against David Sedaris at his best, Calypso, When You Are Engulfed in Flames. Tell me specifically where my sentence-level craft falls short of his. Don't tell me what's good. Tell me which sentences a ruthless editor would cut, which metaphors are clichés I can't see yet, where the rhythm goes flat, and what would be in a Sedaris version that isn't in mine. Rate on hook, scene work, specificity, rhythm, ending, 1 to 10 each, with the reasoning for the grade.
The bamboo pole
Day 23, Patrouille des Glaciers came in highest at 6.2, and only because of one paragraph:
> Kirsten ran around scrambling and the only thing she could find was a broken bamboo pole that had been holding up the course barriers. Great. Six or seven more hours with a bamboo pole. Nice look.
Specific absurd detail. Short punch. Self-implication. Dry wit. Nice look is the cut. I in fact remember laughing about the blue bamboo pole with my ski bro. Gandalf the Blue he was calling me that eve, at the after PdG Barby.
Every BIP post since gets measured against that paragraph. It is the line between MFA-workshop neutral and something I would actually want to read.
The skill got smaller
The first version of the skill tried to do everything. Rate the post 1 to 10. Propose the Sedaris version that isn't in yours. Write 200-word rewrite targets. Two passes in, the skill was rewriting my posts for me. The asset is my voice. The skill's job is to pull out the cheese so the voice gets cleaner, not to swap my voice for Sedaris's.
So the skill shrank. Four scans. Single-phrase replacements only. No paragraph rewrites. No ratings. No what would Sedaris do hypotheticals. Thin-harness, fat-skills applied to a writing skill.
The skill is not doing the writing. The human does the writing. Drafting is the messy human bit. You start by rambling, but contained rambling. You have an idea, you start to type, raw is fine, raw is human, and the machine cannot replicate raw. Even if it lands on the page like Elemental slop (Swiss cheese reference), that's fine. The cleanup happens on the next sweep.
What is in the skill
Four scans, in order.
1. Blacklist scan. Grep the draft against 04-Reference/BIP-Cliche-Blacklist.md. Eight categories, more than seventy specific phrases. The list is growing daily, but here are some samples of what is on the list:
- Beautiful, soul, journey, universe, heartbreaking, powerful, meaningful, epic. Flat-banned as labels applied to a person, a life, or a feeling.
- The darkest stretch. The most common grief phrasing in English.
- Shape, frame, lane. Claude's reflexive structural-metaphor nouns. All three banned in the last two weeks for reading AI-fine. I freaking hate these!!
- It's good to be alive... with the wistful ellipsis. Closer banned after Day 23.
2. Forced simile scan. Any like a X doing Y where the comparison is performing wit. The canonical example, Day 28:
> refreshing the applicant list like a teenager checking a crush's Instagram
Stomach-turning AI-fine. What I changed it to, simple and easy to read:
> I check the applicant list most mornings. Same empty page.
3. Meta-writing scan. This post is not about X. I'll try not to drone. Here are the lessons I'm taking away. That is the whole trick. Here is the part that hurts. Replacement is almost always: cut the sentence. Let the next line stand bare.
4. Abstract closer scan. If the last three words of the post are abstractions (hope, love, peace, meaning), flag it. End on the last concrete image in the prior paragraph instead.
Back to the draft
You have your first draft. Raw, mostly rambles, maybe with a placeholder where you asked the AI to look up a fact you didn't have time to research. The AI co-author is good at that bit. Go fetch the annoying detail facts that would otherwise disrupt the flow. The flow itself is the writer's job. In my case the flow is brain to fingertips, tap, tap, tapping on my keys.
One phrase I remember from my 11th grade English teacher: Omit needless words. Omit needless words. I do not always do that. But I want my AI co-author to think that. Most of the time, if it tells me to take something out, I take it out. Sometimes I override.
Drafting is a separate skill (bip-writing). The Sedaris pass runs after the draft is otherwise ready, before deploy. Output is a list. Each item is one quoted sentence, the cheese type, one proposed replacement. I pick what I feel the skill gets right. I nearly always agree with it. Anything new gets appended to the blacklist with the date and the reason.
Always close the loop.
Cheese counts so far, append-only:
- Day 32: 5 flags
- Day 35: 7 flags
- Day 36: 2 flags
- Day 38: 9 flags
- Day 41: 6 flags (5 applied, 1 left intentional)
- Day 42: 4 flags
Trend is roughly flat, which is honest. Some days I write sharp. Some days I write tired and the AI-fine creeps back in. Daily writing isn't easy and you have to be in the mood for it. Day 36 was a day I was fired up and spent two hours on the post. Will I look back to confirm what that one was about? Maybe.
As an example on time and human input, I wanted to get this one right so I started writing at 6am, and am wrapping it up in between about 5 other tasks at 10:36am.
Why this works for me
Anyone can prompt Claude. Almost nobody writes down what Claude consistently gets wrong about their voice and feeds it back into the next draft. The blacklist plus the corrections log is the only thing standing between my published voice and the average voice of the internet, which is exactly what the model sounds like by default.
The Sedaris skill is the cleanup crew. It is the last thing I run to verify the post is OK, and catch some cheese I'm not aware of, or that the ego overlooks.
I do not get it right every day. But I have a feeling that twelve months of this, daily, will leave me writing in a way I would recognise as more mine and less the model's.
Yesterday's line still applies. AI-fine is the default for everyone now. The work is in the gap between fine and yours.
Monthly Revenues $11,800 | Clients 2 | Prospects fading | Employees: me
Day 43 of 365.