This article shares how AI Coding accelerated development of a Mastodon web client, with a focus on prompt-driven engineering, complex interaction design, and high‑leverage refactors. It highlights prompt patterns, integration strategies, and two key features: React Query and the rich text editor.
Preview: https://v0-mastodon-client.vercel.app
Github: https://github.com/fh332393900/mastodon-client
Model note: Most of the AI work used GPT‑5.2‑Codex.
1) Why AI Coding Here?
The project evolved quickly: rich editor, federated search, emoji handling, code blocks, and multi‑language UI. AI was used as a co‑pilot for high‑impact changes, not just boilerplate.
Key goals:
- Ship complex interactions fast
- Maintain correctness (state boundaries, serialization)
- Keep UI responsive
2) Prompting Patterns That Worked
A) “Diagnose then propose”
Prompt:
“Explain the cause of input lag, then propose a fix with minimal changes.”
Outcome:
AI focused on render priorities and avoided rewriting the component blindly.
B) “Refactor without changing behavior”
Prompt:
“Replace editor implementation but preserve mention/hashtag/emoji output format.”
Outcome:
AI matched existing serialization rules (@user, #tag, :emoji:) and kept backend compatibility.
C) “Complex interaction in one shot”
Prompt:
“Add a toolbar with a popover containing: code block, bold, italic. Each tool highlights based on cursor position.”
Outcome:
AI generated a consistent UI state model and wiring via editor.isActive(...).
3) AI + Rich Text Editor (Key Highlight)
The editor is the most complex UI surface. AI helped:
- Move to Tiptap with custom NodeViews
- Keep mention/hashtag/emoji as atomic nodes
- Export plain text while rendering rich UI
Why it matters:
The editor must feel native while still producing Mastodon‑compatible plain text.
4) AI + React Query (Key Highlight)
React Query powers async data flows (search, timelines, discovery):
- Query‑based caching keeps UI fast
- Stale‑while‑revalidate pattern fits social feeds
- Easy invalidation on user actions
AI benefit:
Fast scaffolding of hooks and query keys with consistent conventions.
5) Lessons Learned
- Be explicit with constraints: “Don’t change serialization output” prevents regressions.
- Split prompts: diagnosis first, refactor later.
- Focus on state boundaries: editing vs. persistence, display vs. submission.
6) Takeaway
AI Coding becomes effective when it’s asked to solve specific engineering problems, not just generate code.
In this project, AI was most valuable for:
✅ Structured refactors
✅ Complex interaction logic
✅ Maintaining behavior while changing architecture