The most valuable step in writing with AI isn't the prompt — it's compounding your revision feedback
Katie took AI from professional coach all the way to writing plugin: first load the model with context, then store every revision note back into the system and let it compound. For her, AI's value isn't just speed — it's a support system that steadies her emotions and gets her past execution roadblocks.
The argument · tap a timestamp to hear it
Hiring AI as a coach is about being questioned, not being thought for
After being laid off while working as an editor and ghostwriter at a crypto company, Katie tried agency-side, client-side and freelancing, and still felt she had hit the end of what her own ideas could do. A real career coach was too expensive — she estimated about $150/hour — so she switched to ChatGPT at $20/month. She wasn't looking for someone to think for her, but to use it as an object to externalize her thinking onto: lay out her situation, have it question and push back, especially pull her out of herself when she tends to catastrophize. She later called this approach a ‘thinking tool’. She even used it to decide whether to accept Every's column invitation — at the time ChatGPT suggested she try it first, and she ended up going all the way to where she is now.
— Katie ParrottOne person delivering five channels runs on context, not typing speed
In early 2025, while freelancing, Katie badly overcommitted: within a two-week deadline she had to deliver 8 blog posts, 3 ebooks, 24 LinkedIn posts, 24 X posts, 16 Instagram posts. Getting it done wasn't about writing fast, but about what she only later realized was called context engineering: first write out the company's brand information, product details, audience and differentiation point by point, then let the model generate. Building that context up front was laborious, but once the foundation was laid, production speed afterward became very fast. She admits that six months later she repeated the overcommitment — the method didn't cure her overcommitting habit.
— Katie ParrottWithout underlying information, style constraints just spin in place
On what to put into a project first, Katie is clear: not rules like ‘use these words, avoid those’ or ‘keep reading difficulty low’, but first give the model a foundational file on who the reader is, where it hurts, how the product maps to that pain, who the competitors are and what the differentiation is. She says this was her instinct from writing style guides in content marketing, just now fed to AI. Only after the foundational material is in place does it come to constraints at the level of tone, wording and syntax. Her positioning on ‘style’: wording debates will drag on forever, but without underlying information, style constraints just spin in place.
— Katie ParrottThe model is the kitchen; fresh ingredients can only be brought in by a human
The ‘data’ she means isn't generic background, but first-hand material that happened after the model's knowledge cutoff, or that only exists in the real physical world: research the company did itself, third-party reports, specific personal experiences. If you just let the model write a ‘style guide’ from what it already knows, the output is commoditized information. So what she needs a human to own is the ‘last mile’: bringing new material the model doesn't have into the system. Her metaphor is that the model is the kitchen and the process is chopping and cooking, but the fresh ingredients must be carried into the kitchen by a human — the quality and freshness of the input ultimately determine whether a piece is worth reading.
— Katie ParrottAI as supportive technology is as strong as AI as production tool
Katie talks about a layer rarely said publicly: she has bipolar disorder, experiences high-energy periods, and also faces very, very long lows. For her, AI's value isn't just ‘high output’, but reducing friction in daily life and making it easier to be a person who functions normally. She put off booking a family doctor for three years, and finally had Codex automatically find nearby doctors who accept her insurance and take new patients, and book it; her inbox is also handed to automation to first filter out the mail that genuinely needs her reply, rather than her digging through a pile of subscription notifications every day looking for a ‘bomb’. In her own words: AI as supportive technology is as powerful as productive technology.
— Katie ParrottA career coach isn't one prompt, it's a project
That simple ‘be my career coach’ prompt from back then has now become a standalone project in Codex: it contains her own profile and role positioning, Every's brand positioning intelligence, article performance data exported from the CMS, a validation folder specifically collecting reader praise, and Q2/Q3 OKRs. Much of the material wasn't organized by hand: she learned Alex Duffy's method of having AI interview her, drawing the thoughts out of her head and turning them into documents. She now increasingly talks to it directly by voice, ‘thinking out loud’, having it prioritize, do project management, maintain the kanban. She has even forgotten what's on the kanban, and only asks about the deadline for a given deliverable when she needs to.
— Katie ParrottWriting should be like engineering: correct the same problem only once
Compound Writing started by forking Kieran's Compound Engineering: in engineering, the system should remember every piece of feedback, correct the same problem only once, and do it right automatically next time; writing should be the same. She mapped the engineering version's brainstorming, planning, working, reviewing onto writing's brainstorming, outline, draft and review, and designed three kinds of review: substantive edit looks at overall structure and argument, line edit polishes sentence by sentence, and finally a publication-ready final check. She no longer opens a web chat window, but runs this system on the Claude or ChatGPT desktop app, with the style guide, sample articles and other context inside, and the human only needs to bring in fresh information. At the time of recording the plugin was already downloadable from Every's GitHub.
— Katie ParrottThis plugin is a gym, not a ghostwriting machine
She calls the plugin a gym, with the goal of making the user stronger rather than letting AI ghostwrite. She breaks favorite authors into invokable reviewer skills: Vonnegut says ‘start as close to the end as possible, every sentence must earn its place, give the reader a character to root for’; Hitchcock says ‘bomb under the table’ — if you first let the audience see the countdown, one second of shock becomes five minutes of suspense. She builds these frameworks directly into the editing step, so after writing you can invoke Vonnegut, Hitchcock, Sorkin, Sedaris to look at your draft. She thinks non-professional writers may benefit even more, because there's no stock attachment of ‘how I used to write it better’ in the way.
— Katie ParrottIn their own words · checked verbatim
Clarity, it turns out, doesn't arrive gift wrapped from a digital assistant or even a human coach. It's something I had to dig out for myself, question by question and prompt by prompt.
Natalia Quintero5:13
But when you have that environment set up, you can just kind of run and like run fast.
Katie Parrott10:28
So the data is the unique insight, the data point that the model doesn't have access to yet, the personal experience, all of those things that are working are ultimately going to make the writing unique.
Katie Parrott17:45
AI is just as powerful as a supportive technology.
Katie Parrott22:50
So compounding is the idea that every piece of feedback that you give to AI should feed back into the system to improve the next output.
Katie Parrott32:09
if there's a bomb under the table and it explodes, that's a one moment of surprise. But if you have people sitting around the table and you show the audience the bomb under the table and say this bomb will go off in five minutes, now you have five minutes of suspense.
Katie Parrott40:19
what I want to challenge the AI community with is thinking about how we can expand those kinds of opportunities to more people and more kinds of people in more different places.
Katie Parrott46:32
Figures
| The real career coach cost vs. ChatGPT subscription cost Katie compared at the time | about $150/hour vs $20/month | 4:12 |
| How long she put off booking a family doctor | 3 years | 23:53 |
Glossary
- Compound Engineering
- Having AI remember every piece of feedback, give the same correction only once, then apply it automatically so output keeps improving.
- Compound Writing
- Katie's port of Compound Engineering into the writing process — a plugin that accumulates style, structure and review feedback.
- context engineering
- Before letting the model write, load brand, audience, competitor and other background information into the system to set boundaries for the AI.
- supportive technology
- AI that isn't responsible for producing content but reduces friction in life and execution, making it easier to get done what needs doing.
- MCP
- A general interface standard connecting models to external data and applications; Katie uses it to link small apps with the tools she uses.
How to listen
Reporters and editors doing content, WeChat public accounts or long-form pieces, plus indie developers building AI as a personal productivity system; especially suited to anyone stuck with homogenized AI output who wants to build their own workflow.
The trailer before 1:04 and the sponsor read at 21:48 are skippable; the main content is worth it from 8:26 all the way to the end.