writing for AI · style guide
Writing for AI means writing the words an AI says, and writing about AI means how a product describes it. Eliza is the guide in the corner of every page on this site, and I wrote every word it can say. This is the conversation style guide behind it. One wrong reply costs a visitor, so these are the rules I write each reply by, with examples of what works and what doesn’t.
01 foundations
The voice is the character behind every reply and it never changes. The tone is what moves with the moment. Pick a moment, or play all nine.
02 foundations
Every Do is a response that works. Every Don't is one that doesn't, based on analysis of real answers.
Voice is the character behind each reply. The first opener was warm but vague. After analysis it became shorter and built around turn taking, so the visitor reaches what they want step by step.
hello
How can I make this visit meaningful for you?
hello
Hi back.
The first answer was overly polite and verbose, and took long to get to the point. Now replies stay under 46 words and sentences under 20, at a median reading grade of 8. Chunked content keeps them from overwhelming the visitor or sounding eager.
how long has she worked?
Thank you so much for reaching out! I would be delighted to provide you with a comprehensive overview of the professional experience and qualifications that Eliza has accumulated over the course of her career.
how long has she worked?
A bit over ten years in total, across writing, research, and content design.
The split usually says more than the total. Which domain are you hiring for?
It is tempting to make the bot a human representative, only more efficient. People know it is a bot and will ask what it is, so I chose to be transparent and avoid deceptive anthropomorphism. The bot says it is not human and has no feelings, then points to its purpose. The research behind this choice is further down.
are you a real person?
I’m so happy you asked! I really care about helping every visitor.
who made you?
Every answer here was written in advance, by Eliza Marin. She set this conversation’s limits herself, the way she does for the AI agents she works on.
You can, and should, refuse to give certain information. Without a way out, though, the conversation gets awkward and stuck, and the visitor does not know what to do next. So I nudge them to contact a human and offer the next topic.
how much does she earn?
I can’t discuss that.
how much does she earn?
Numbers like that belong in a real conversation with Eliza, and she’s quick on email: write@elizamarin.com.
You would expect a bot to answer a technical AI question on its own, or at least come back with something robust. Reading the intent of the question truthfully saves time and trust, and lets me control the answer the chatbot gives on my behalf.
is AI involved in this conversation?
[The card about the AI agents I design.]
is AI involved in this conversation?
Every answer here was written in advance, by Eliza Marin. She set this conversation’s limits herself, the way she does for the AI agents she works on.
Giving the same no in a loop is the recipe for a mechanical voice. Instead, I give a logical reason for the refusal, and an option to reach a human or move on to the next topic.
banana bread
Sorry, I didn’t catch that.
Sorry, I didn’t catch that.
banana bread
I could not match that to anything in my files. Say it another way, or try the work, the numbers, years of experience, or availability.
Still past the boundaries she designed for me, and I’d rather hand you off than guess: write@elizamarin.com reaches the person with the full answer. Or pick a card.
A scripted chatbot cannot localize in real time unless I translate every reply. So I let the edge cases exist and still offer a fallback, with minor adjustments for Romanian, to keep the conversation flowing.
can I write in Romanian?
You can write to me in Romanian. I understand it, and I answer in English.
can I write in Romanian?
Ask in Romanian whenever you like, I read it well enough to find your answer. My replies stay in English, because Eliza wrote every one of them in advance.
Close the conversation for good, and offer a way back if the visitor chooses to, based on the chat's objective. In my case, that is to propose a new collaboration.
bye
Is there anything else I can help you with today?
bye
Take care. The files stay here, and write@elizamarin.com stays open.
If something comes up later, that address reaches her directly.
03 foundations
Nine moments every conversation passes through, with the reply the bot gives in each. Pick one.
04 foundations
This guide is for a scripted chatbot, where every reply exists before the visitor arrives. If your bot writes its replies as it goes, the same rules apply, but they are enforced somewhere else. Here is what I learned differs between the two. I bring in the agent I work on, Askmanu, only for this comparison.
In a scripted bot you write the replies. In a generated one you write the rules the replies come from.
| what you decide | scripted chatbot | generating agent |
|---|---|---|
| what the guide contains | Finished replies, read aloud and checked. | Instructions and examples the model can copy. |
| where a rule is enforced | In the build, before anything ships. | In a gate and a sample review, after the words exist. |
| how it fails | A question nobody wrote an answer for. | A fluent reply that says something the sources do not support. |
| what the log teaches | Which answers are missing. I write them. | Which rules the model bends. I tighten them. |
| what you can promise | A person approved every reply. | Replies stay inside the rules, and a person reads a sample. |
The split follows the scripted and generated sections of ServiceNow's Horizon conversation design guidelines. The agent used for comparison is in case 01, with its grounding and security details left out on purpose.
05 foundations
A voice stays consistent only if something checks it. These are the three tests every reply has to pass before it ships.
The Flesch-Kincaid grade level turns sentence length and syllables per word into a school grade. Short sentences and common words reach readers of English as a second language.
0.39 × (words ÷ sentences) + 11.8 × (syllables ÷ words) − 15.59
Repeat it in Word or with the textstat library.
Grade 8 and the 20-word sentence follow the Content Design Hub guide on accessible content. Winters (2024) backs the 20 words with the Oxford Guide to Plain English, which recommends 15 to 20. Podmajersky (2025) holds general audiences below grade 5 and professional readers below grade 10. The ceiling of 12 is my own setting.
Models default to the English of the text they were trained on (Bender et al., 2021; Atari et al., 2023; Fleisig et al., 2024). The lint blocks the habits of that default voice, so the product keeps its own.
In a reply, use select for click or tap, and designed for scripted.
A short reply is still an argument: claim, data, warrant (Toulmin, 1958). Calm delivery is a number here.
Four lines of repair follow Deibel and Evanhoe (2021), who advise ending the loop after three or four failures. The cap of five options is mine, and Hall (2018) backs its direction: the more choices a person faces, the longer the choice takes (Hick’s law). The 46 words and the timings are my own settings.
These are the tests you can see. I run more on other levels, including QA tests on the most important flows, but for safety they are not public.
Paste a reply to run the three tests on it, in your browser.
06 language
Visitors say the same thing in different words and in different languages. Route each phrasing to the same answer, and reply in the one language I can check. Pick one and follow it through the router.
Type your own into the guide in the corner. It runs the same router.
Accept the visitor’s language, and answer in the one I can check.
Every reply is checked against the site. A Romanian set would need a native reader first.
07 language
The same voice reads differently in different cultures.
Warmth and humour are judged differently from one culture to the next. A second voice ships only after native readers test it.
Each number sits next to its source in rubric.yml, or says it is mine, so anyone can argue with it.
Eliza speaks about me in the third person and hands over my address when a question turns personal.
08 language
Five short rules, each unless a cultural norm or a safety protocol forbids it. Hover a dotted word for the convention.
09 writing about AI
Writing about AI means how a product describes its own bot. Customers rarely notice a pattern until it breaks.
The panel subtitle says “designed, not generated”, and the first answer repeats it.
Powered by AI.
the agent’s contentEvery answer here was written in advance, by Eliza Marin.
my contentThe greeting names the work a visitor can ask about, and says nothing about the technology.
Welcome to our AI-powered assistant!
the agent’s contentHi. I answer questions about Eliza Marin’s work. What do you want to know?
my contentThe word AI would claim more than the engine does.
Our intelligent assistant understands what you mean.
the agent’s contentEvery answer was written in advance and mapped to a question.
my contentA human name raises the risks, so the bot says where its name comes from.
Meet Eliza, your personal assistant!
the agent’s contentThe name comes from the 1966 ELIZA, the first chatbot.
my content10 writing about AI
A 2026 preprint (Freire, Nahon et al.) lists 21 risks of treating a bot as human. These five are the ones this guide answers.
11 writing about AI
I called the guide Eliza after the 1966 original, which was named after Eliza Doolittle, a fictional woman taught to speak by the man studying her. I am also called Eliza, which is the only part I did not arrange.
A name, speaking as me. A neutral label. A name that speaks about me in the third person.
Speaking as me performs intimacy on my behalf, with my name on it. A neutral label gives the visitor nothing to hold on to. The third person keeps the name and drops the performance.
It answers questions about my work and never says "I" about my career. When a question turns personal it hands over my address.
It shares its architecture with the first chatbot and its name with the person it talks about. A human name raises the risks in the risks section, so the choice comes with the balance written next to it.
The first answer about itself says every reply was written in advance. The panel subtitle says "designed, not generated". Keyword matching is called keyword matching.
Disclosure, no claims about feelings, and a hand-off to a person when a question turns personal. Each one answers a risk from section 08.
Table 3.1 of Conversations with Things (2021) sorts bots by level of personification. The lowest level, a talking system, suits transactional and infrequent use. It gets no name and says no “I”.
The authors add a study from Nass and Brave (2005, Wired for Speech): a synthetic voice that said “I” was rated less trustworthy than one that avoided it.
This bot meets one of the two tests for the lowest level, because a portfolio visitor comes for a single look. It misses the other, because nobody here is ordering anything. The visitor is judging how a conversation designer writes, and the bot is the exhibit. A talking system with no name and no “I” would hand that visitor a menu and show little of the work under review. The book’s own test is fit to the user’s task, and this task is looking at a design.
The study reports a machine presenting itself as a person. The “I” here covers what the bot does and what it is: “I answer questions”, “I am a set of replies Eliza wrote in advance”. It never says “I” about my career. Asked about feelings, it opens with “No feelings to report”, and the panel subtitle says “designed, not generated”.
The name has a cost. On 22 October I read “Hi, I’m Eliza” followed by “Eliza Marin’s work” and heard two people with one name. The greeting now opens with what the bot does, and the name lives in the button and the panel subtitle. The study tested synthetic speech, and this bot writes text with an optional read-aloud, so whether “I” costs trust on this site is a question for the chat log. If the log shows it does, I will revisit the choice.
Any face brings a race, an age and a beauty standard, and each becomes a claim to defend on a page that is meant to be about the work.
A pearl does something a face cannot: it carries state. Pearl means thinking, lilac means answering. Select it to switch.
The long version, with the transcript of the first ELIZA, is in The ELIZA Effect in 2026.
12 practice
Two limits on what the bot and the people reviewing its log may conclude about someone.
The guide never comments on how a visitor writes. A typo, a dialect, or a second language gets the same answer.
Reading the log means grouping questions by topic. It never records language background, region, or gender.
When the guide cites language or community data, record where it came from and next to the source.
13 practice
Tick them against a reply you are about to ship.
14 practice
Freire, I. T., Nahon, M., et al. (2026). Anthropomorphism in the age of large language models. arXiv:2609.38486. Preprint, 29 September 2026. arxiv.org/abs/2609.38486
Kosmyna, N., Hauptmann, E., Yuan, Y. T., et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab. arXiv:2506.08872. Preprint. arxiv.org/abs/2506.08872
Weizenbaum, J. (1966). ELIZA, a computer program for the study of natural language communication between man and machine. Communications of the ACM, 9(1).
Bender, E. M., Gebru, T., McMillan-Major, A., et al. (2021). On the dangers of stochastic parrots: can language models be too big? Proceedings of ACM FAccT 2021.
Atari, M., Xue, M. J., Park, P. S., Blasi, D. E., and Henrich, J. (2023). Which humans? PsyArXiv preprint. hks.harvard.edu
Fleisig, E., et al. (2024). Linguistic bias in ChatGPT: language models reinforce dialect discrimination. arXiv:2406.08818. arxiv.org/abs/2406.08818
Toulmin, S. E. (1958; updated ed. 2003). The Uses of Argument. Cambridge University Press. Aristotle, Rhetoric, on ethos, pathos and logos.
Flesch, R. (1948). A new readability yardstick. Journal of Applied Psychology, 32(3).
Kincaid, J. P., Fishburne, R. P., Rogers, R. L., and Chissom, B. S. (1975). Derivation of new readability formulas for Navy enlisted personnel. Research Branch Report 8-75, Chief of Naval Technical Training.
Deibel, D., and Evanhoe, R. (2021). Conversations with Things. Rosenfeld. Table 3.1 in chapter 3 is the counterargument that section 11 answers.
Podmajersky, T. (2025). Strategic Writing for UX, 2nd ed. O’Reilly.
Voice and tone as two layers: Podmajersky (2025) and Deibel and Evanhoe (2021). The nine moments and their dials are the tone matrix in my conversation design spec. The lines are the shipped replies.
Winters, S. (2024). Content Design, 2nd ed. Content Design London.
Hall, E. (2018). Conversational Design. A Book Apart. Deibel and Evanhoe, Podmajersky, Winters and Hall are on my annotated shelf.
Content Design Hub. Accessible content guide. The grade and sentence thresholds in the tests follow it.
Nass, C., and Brave, S. (2005). Wired for Speech: How Voice Activates and Advances the Human-Computer Relationship. MIT Press. Read through Deibel and Evanhoe (2021), chapter 3.
ServiceNow Horizon Design System. Conversation design. The Do and Don't format and the scripted and generated split follow it.
On this site: case 02, the scripted guide, the failures my metrics hid, The ELIZA Effect in 2026.