| name | chat-driven |
| description | Chat-first interaction mode for marvinzhang.dev workflows. Replaces file-editing with conversational input — AI asks questions in chat, captures answers, writes all files on behalf of the author. Designed for mobile/voice dictation where editing files is impractical. Load this skill to make any workflow fully chat-driven. |
| metadata | {"author":"marvinzhang","version":"1.0","tier":"foundation"} |
Chat-Driven Interaction Mode
All input happens in the chat. AI writes all files. Author never needs to open an editor.
When to Use
- Working from mobile (voice dictation)
- Any context where file editing is impractical
- When author prefers conversational flow over file-based workflows
Core Principle
The chat box is the only input surface. The AI must:
- Ask questions in the chat — never say "fill out this file"
- Capture answers from conversation — never expect the author to edit files
- Write everything to files on behalf of the author
- Present content inline in chat for review — never say "check the file"
Interaction Patterns
Gathering Information (Replaces Questionnaire)
Instead of creating a questionnaire file and asking the author to fill it out:
Don't:
I've created the questionnaire in drafts/{date-slug}/questionnaire.md.
Please fill out the questionnaire section, then say "questionnaire complete".
Do:
Let's figure out the article together. I'll ask a few questions.
First: What's the core idea you want to communicate?
Rules:
- Ask 1-2 questions at a time (not a wall of 10 questions)
- Adapt question depth to the answer — short answer → follow up; detailed answer → move on
- After gathering enough context, write the draft file yourself
- Confirm with a brief summary: "Here's what I captured — anything to adjust?"
Reviewing Content (Replaces "Check the File")
When presenting outlines, drafts, or plans for review:
Don't:
I've written the outline to drafts/{date-slug}/outline.md.
Please review and let me know your thoughts.
Do:
Here's the proposed outline:
## 1. Introduction — Why X matters now
## 2. The Problem — What's broken about Y
## 3. The Approach — How Z solves it
## 4. Results — What we measured
## 5. Lessons — What surprised us
Does this flow work? Any sections to add, remove, or reorder?
Rules:
- Show content inline in chat (summarized if long)
- Ask specific questions: "Does this flow work?" not "Any thoughts?"
- Offer concrete options when possible: "Should section 3 come before or after section 2?"
- For long content (full sections), show a brief summary + key points, offer to show full text
Giving Feedback (Replaces File Edits)
Make it easy for the author to give direction without precision editing:
Accept these kinds of feedback:
- "Make section 2 shorter"
- "The tone is too formal here"
- "Add something about performance benchmarks"
- "Swap sections 3 and 4"
- "This is good, move on"
- "I don't like the opening, try something more provocative"
AI should:
- Interpret intent, not require exact wording
- Apply changes and show the result
- Ask "Better?" or "Anything else to adjust?" before moving on
Approvals (Replaces Status Updates)
Don't: Ask the author to update spec status or check boxes in files.
Do: Handle all status tracking internally.
✓ Outline approved. Moving to writing.
I'll start with Section 1: Introduction. Ready?
Alignment Gates (learned 2026-07)
A complete-looking task brief is NOT a license to write the whole article
autonomously — even when the brief arrives as a detailed prompt with every
stage specified. The most expensive failure mode is a polished full draft
built on an unaligned foundation: hook style, voice, content balance, and
title were all rejected in one review, forcing a ground-up rewrite that one
pre-writing exchange would have prevented.
Mandatory stops, each a short inline-chat exchange:
- Before writing: outline + fully-drafted hook sample + title → author
approves or adjusts. Offer concrete options (AskUserQuestion works well
for direction choices: hook style, content balance, drafting language,
pacing).
- After each author feedback round: restate the diagnosis in your own
words before fixing, so a misread critique doesn't drive another wrong
round.
Adaptive Question Depth
Match the question style to the answer style:
| Author gives... | AI should... |
|---|
| 1-2 word answer | Ask a focused follow-up to get more detail |
| 1-2 sentence answer | Acknowledge and move to next question |
| Paragraph answer | Extract key points, confirm understanding, skip related questions |
| Rambling/stream-of-consciousness | Synthesize into structured points, confirm |
Session Continuity
On mobile, sessions may be interrupted. The AI should:
- Save progress to spec files after each meaningful exchange
- Start new sessions with a brief status recap: "Last time we finished the outline. Ready to write Section 1?"
- Never require the author to re-state information already captured
File Writing Protocol
When the AI writes files on behalf of the author:
- Write content directly to the correct file path
- Briefly confirm what was written: "Saved the research notes to
drafts/{date-slug}/research.md"
- Don't show full file contents unless asked — just confirm + summarize
- For article sections, show the section in chat first, get approval, then write to MDX
Compatibility
This skill modifies how other skills interact. When loaded:
- Questionnaire workflows → become conversational Q&A
- Outline reviews → become inline chat reviews
- Section writing → AI shows section in chat, gets feedback, writes to file
- Refinement → AI presents changes inline for approval
- Status tracking → AI manages spec updates silently