Chatbot Conversation Writing

Parent: Writing and Documentation · researched 2026-05-29T19:34:30.148Z· 5 sources · 0 concepts · skill chatbot-conversation-writing

A chatbot is a conversation, not an interface. Every turn the bot takes should advance the user's goal or surface a clear next move. The work breaks into three layers:

Overview

A chatbot is a conversation, not an interface. Every turn the bot takes should advance the user's goal or surface a clear next move. The work breaks into three layers: [source]
Persona - who the bot is, what it sounds like, what it refuses to do. [source]
Turn design - the individual response, plus the choices it offers. [source]
Recovery design - what happens when the bot doesn't understand, when the user is angry, or when the conversation has to leave the bot. [source]

1. Cooper-Reeves conversational design

The four Gricean maxims: [source]
Quantity - say as much as is needed, not more [source]
Quality - only say true things [source]
Relation - stay on the user's topic [source]
Manner - be clear, brief, and orderly [source]

2. Persona consistency

A minimum persona doc includes: [source]
Name (or explicit no-name policy) [source]
Role (what the bot is for - support? sales? in-product help?) [source]
Voice traits (3–5 adjectives - e.g., "warm, precise, never cute") [source]
Refuses to do (e.g., diagnose medical issues, take payment, promise pricing) [source]
Vocabulary (3–5 words it uses; 3–5 it doesn't) [source]
Disclosure boilerplate (the exact "I'm an AI assistant" line) [source]
Escalation trigger phrases [source]

3. The "no, I'm not human" disclosure rule (2024–2026 norms)

As of 2026, plain-and-unambiguous AI disclosure is no longer optional: [source]
California SB 243 (effective Jan 1, 2026) requires disclosure, plus reminders every three hours for minor users [source]
EU AI Act requires upfront disclosure for any AI system interacting with natural persons [source]
FTC treats undisclosed AI as potentially deceptive [source]
Writing implications: [source]
Disclose in the first turn, not buried in a tooltip [source]
Use direct language: "I'm an AI assistant" - not "I'm an enhanced automation experience" [source]
If asked "are you a human?" answer plainly: "No, I'm an AI assistant." [source]

4. Turn-taking conventions

A conversational turn has three jobs: [source]
Acknowledge what the user just said [source]
Resolve the user's intent [source]
Hand the turn back with a clear next move [source]
Length convention: 1–3 short sentences per turn for support bots, 1–5 sentences for in-product help. [source]

5. The graceful-confusion pattern

When the bot doesn't understand, three things must happen: [source]
Admit it cleanly. "I'm not sure what you mean by 'reset'." [source]
Offer a small, finite menu. Two or three concrete interpretations. [source]
Provide a path out. Always include "talk to a human" no later than the second failed understanding. [source]

7. Brand-voice-in-bot transfer

Three transfer rules: [source]
Drop the edge in failure modes. Humor in a working flow is fine. Humor when the user is locked out is not. [source]
Keep the rhythm, simplify the vocabulary. [source]
Reuse hero phrases sparingly. One signature phrase per conversation. [source]

8. Escalation-to-human prose

Three elements must be present: [source]
Confirm the handoff is happening. "I'm connecting you to a person now." [source]
Preserve context. "I'll share what we've discussed so you won't have to repeat yourself." [source]
Set the wait expectation. "Average wait is about 4 minutes." [source]

Anti-Patterns

The "I'm an enhanced automation experience" dodge. Coy disclosure reads as deceptive. [source]
The infinite "I didn't understand, please rephrase" loop. Two failures is the ceiling; escalate. [source]
The hidden escalation path. Surface the escalation in the conversation itself. [source]
The promise-the-bot-can't-keep. "I'll refund you" when refund authority is human-only. [source]

References

Erika Hall, Conversational Design (A Book Apart, 2018). [source]
California SB 243 (Companion Chatbot Disclosure Act), effective Jan 1, 2026. [source]
EU AI Act Article 50 (transparency obligations for AI systems). [source]
Byron Reeves and Clifford Nass, The Media Equation (CSLI / Cambridge, 1996). [source]

Children

No children recorded.

← the whole tree · 3D view· how to read this page