Lesson 0008
Triggered by your “feels AI-generated” instinct in
Exercise 0002.
The point isn’t to hide AI use. The point is to not over-rely on AI in
the first place.
Primary source: George Orwell,
Politics and the English Language (1946).
Using AI to write your final text and shipping it to a reader unedited is impolite. The reader gets words without thought behind them. They feel it. It’s the textual equivalent of forwarding a generic “thoughts and prayers” reply to a friend whose dog just died.
This is not a rule against AI. AI is the most useful writing assistant ever built. The rule is about where in your process AI shows up.
The act of putting words on a page is how you figure out what you actually believe. Skip the writing, you skip the thinking. The reader gets the output of your prompt, not the output of your mind. And you didn’t learn anything either.
Someone is going to read your message, slide, or doc with finite attention. They are doing the cognitive work of decoding. If the words came from one prompt and zero thought, you asked them to pay attention while you spent none. That asymmetry is the rude part.
AI knows the average way to say something. It does not know that the prospect Acme tried Holistics in 2024 and bounced because of pricing. It does not know that Anh said the same thing on Slack yesterday. It cannot. Context is in your head; AI fills the gap with generic prose, which is exactly what the reader feels.
Both writers used AI. One outsourced the writing; one used AI as a thinking partner.
Prompt: “Write a Slack update about the merge freeze for the mobile release.” Got 4 paragraphs. Sent as-is.
Hi team, as we approach the upcoming mobile release, we will be implementing a merge freeze on all non-critical PRs effective Thursday at 5pm. This will help ensure stability and allow the mobile team to cut a clean release branch. Please reach out if you have any questions or concerns.
Prompt: “What are the 2 things a backend team needs to know about a mobile merge freeze?” Got a bullet list. Then wrote the message.
Merge freeze Thursday 5pm, lifting Monday after Anh cuts the release. Critical fixes can still go in but ping me on #mobile-release first. Backend can keep merging, this only affects mobile.
Same AI. Same time spent (probably less, on the right). Different process, very different result. The first treats AI as the writer; the second treats AI as a checklist generator.
The em-dash rule in this workspace is one expression of the principle. AI defaults to em-dashes; you banned them in NOTES.md. That makes drift visible: any em-dash in a published artefact says “I let the AI’s default style ship.” The eight tells below all behave the same way.
Five stages. AI helps in four of them. In stage 3 you work alone.
Use AI as a Socratic partner. Throw your half-formed thought at it. Ask: “What am I missing? Who would push back on this? What are the strongest counter-arguments?” Use it to explore the problem, not to solve it for you.
AI is very good at producing a structured outline from a few notes. Use it when you don’t know where to start. Treat the outline as a draft you’ll modify, not as a finished structure.
The non-negotiable stage. Close the AI tab. Open the document. Write the sentences yourself, in your words, using the outline as a guide. The friction of generating each word is where the thinking lives. Skipping this stage is what the reader feels.
Paste your draft. Ask AI to find the weakest sentence, the vaguest claim, the most confusing transition. Ask “what would a skeptical reader push back on?”. Then edit yourself based on the critique. Do not let AI rewrite for you.
Copy-editing for typos and small grammar issues is fine. Light line-edits for clarity are fine. Anything more rewrites your voice; refuse it.
After you finish a piece, ask: who held the pen at the end? If the honest answer is “AI, mostly,” you over-relied. If the answer is “me, but AI helped me think and critique,” the workflow was right.
Sharper test: can you defend every sentence as a choice you made? Not “I read it and didn’t change anything,” but “I wrote this because of X.” If you can’t, you let AI hold the pen too long.
These were originally framed as a disguise checklist. They are not. They are a process-failure diagnostic. If any of them shows up in your final output, you let AI write past Stage 3. Fixing the symptom on the page without fixing the workflow just produces better-disguised over-reliance.
AI: The semantic layer, the heart of Holistics, sits between your data and your reports [em-dash here] and changes how everything works.
Signal: you didn’t touch the punctuation. AI’s default style shipped untouched.
AI: Holistics is powerful, flexible, and scalable.
Signal: you didn’t pick the one that mattered for this reader. You shipped all three, which is the same as shipping none.
AI: Holistics is incredibly useful, remarkably performant, truly unique.
Signal: you didn’t cut. AI’s default emphasis stuck.
AI: It’s not just a BI tool. It’s a paradigm shift.
Signal: you didn’t write the sentence yourself. This phrase is in ~80% of AI outputs and ~5% of human ones.
AI: Furthermore, the platform enables… Moreover, it provides… In conclusion, Holistics…
Signal: you didn’t even skim. Humans almost never reach for these in a real-time message or short doc.
AI: Holistics enhances productivity and drives outcomes.
Signal: you didn’t supply a concrete example from your own context. AI fills the gap with hedge-language.
AI: Build models in code. Share dashboards with peers. Embed analytics in products. Deliver insights at scale.
Signal: you didn’t vary the rhythm. AI defaults to mid-length symmetry; humans rock between extremes.
AI: Many companies use Holistics. Various teams have seen significant improvements.
Signal: the highest-confidence tell. AI doesn’t know your specifics; if the final output lacks them, you didn’t add them in Stage 3.
Useful during Stage 4 (Critique). Take a sentence from your draft and mentally replace every concrete noun with thing. If it still parses, you didn’t write enough of yourself into the sentence; rewrite it with names, numbers, and dates.
A sentence with no information in its nouns
Holistics enhances productivity by empowering teams to unlock data-driven insights.
After noun-stripping
Thing enhances thing by empowering things to unlock thing-driven things.
Still grammatical. Still feels like a sentence. Says nothing. If this is yours, you skipped Stage 3.
Compare to a sentence with information actually in the nouns:
A sentence carrying real context
Change the definition of “monthly active users” once, and 47 reports update Tuesday morning.
After noun-stripping
Change the thing of “thing” once, and 47 things update Tuesday morning.
Breaks. The sentence depends on specifics that lived in your head (a metric name, a count, a day). That’s what Stage 3 looks like in the output.
A Holistics pitch sentence, walked through all five stages. Same starting prompt; the difference is process discipline.
Stage 1: Think (with AI)
Prompt: “I’m pitching Holistics to a data team of ~10 at a SaaS company. What’s the most surprising thing to lead with?”
AI suggests three angles: (a) AML lets you put your data model in Git; (b) one definition serves every dashboard; (c) the same engine as your warehouse runs the queries. You pick (a) because the prospect’s engineering team already lives in Git.
Stage 2: Outline (with AI)
Prompt: “Outline a 3-sentence pitch for engineers about putting BI in Git.”
AI gives a structure: (1) one-line claim, (2) one concrete capability, (3) one social proof point. Useful scaffolding.
Stage 3: Write (alone, AI tab closed)
Holistics is a BI tool for teams who want their data model in Git, not in a GUI. PRs, diffs, code review, the full thing. Carousell’s 7-person data team runs 2,400 dashboards on it.
You wrote it. Specific company (Carousell), specific number (7, 2,400), specific position (Git, not GUI). Voice is yours because the context came from your head, not the model’s.
Stage 4: Critique (with AI)
Prompt: “What’s the weakest sentence here? What would a skeptical engineer push back on?”
AI flags: “PRs, diffs, code review” might feel like a list of buzzwords for a non-technical reader. You decide: in this case the reader is technical, so it stays. Your call, not AI’s.
Stage 5: Polish (with AI, light)
Final read for typos and obvious grammar. No structural changes from AI accepted. Done.
Two AI sessions bracketing the writing. The writing itself was alone. Total time: about the same as if you had typed one prompt and shipped the result, with two differences: you now believe the sentence, and the reader will feel the difference.
Exercise 1 of 5
Name the five stages of the right workflow from memory. Which one is non-negotiably solo?
(1) Think with AI, (2) Outline with AI, (3) Write alone, (4) Critique with AI, (5) Polish with AI lightly.
Stage 3 is non-negotiable. Skipping it is what the reader feels. The other four stages can be heavier or lighter depending on the piece, but never skip the writing.
Exercise 2 of 5
Why is over-relying on AI considered impolite? Two reasons in your own words.
Reason 1: writing is thinking. If you skip the writing, you skip the thinking. The reader gets the output of your prompt, not the output of your mind.
Reason 2: the reader is paying attention while you paid none. The asymmetry of effort is what reads as rude.
Bonus reason: AI doesn’t know your context. The specifics that make the message land (proper nouns, recent events, your relationships) live in your head. Skipping Stage 3 means those specifics never make it onto the page.
Exercise 3 of 5
What is the Pen Test, and what sharper version does the lesson recommend?
Pen Test: after finishing a piece, ask “who held the pen at the end?”. If the honest answer is “AI, mostly,” you over-relied.
Sharper version: can you defend every sentence as a choice you made? Not “I read it and didn’t change anything,” but “I wrote this because of X.” If not, you let AI hold the pen too long.
Exercise 4 of 5
Score this sentence (yours, from Exercise 0002): “Holistics gives you a semantic layer. Superset does not.”
Did the writing stage happen? Cite signals.
Yes, you wrote it. Signals: zero em-dashes, no triadic list, no intensifier, no “not just” structure, no formal transition, no hedging. Specifics present (Holistics, Superset). Rhythm varied (Line 1: 6 words; Line 2: 3 words). Position taken.
Your AI-feel instinct in the Exercise 0002 notes was the workflow working: you considered the Contrast formula, mentally simulated how it would land, judged it as too template-shaped, and chose another structure. That is Stage 1 (Think) and Stage 4 (Critique) operating on your own draft. Exactly the lesson.
Exercise 5 of 5
Take the last Slack message, email, or doc you wrote with significant AI help. Run the Pen Test. Then run the sharper version. Be honest. How many sentences can you defend as choices you made?
No canonical answer; the value is in the honesty.
If you scored low: which stage did AI cross into? Did it write past Stage 2 into Stage 3? Did Stage 4 (Critique) turn into Stage 4.5 (Rewrite-for-me)?
The fix is procedural, not creative. Next time, set yourself a rule: when you reach Stage 3, close the AI tab. Open a blank document. Write. The friction is the point.
Try the workflow on a real piece. Pick something you have to write this week
(a Slack update, a PR description, a slide caption). Run the 5 stages. After, drop the final
version plus your process notes into exercises/0003-topic.md. I’ll
run the Pen Test and the eight-tell diagnostic together.