Depend on typed provider ports
Depend on typed provider ports instead of provider response shapes, making fallback replaceable at adapter boundaries.
Thoughtline - LinkedIn Writing Assistant
extension
User-controlled Chrome side-panel extension for drafting replies from pasted or selected LinkedIn posts, refining writing, and keeping every publish action manual.
- Role
- Product design and browser-extension engineering
- Scope
- Browser integration, provider workflows, and privacy
- Status
- GitHub release available
Problem
Writing a useful LinkedIn reply or post requires understanding the visible conversation while preserving the author's own voice, intent, and control.
Constraints
- C1
LinkedIn page content is untrusted and must be bounded before entering prompts.
- C2
OpenRouter, Gemini, and Groq keys are all required so the fixed free-only route can fall back without asking during a writing task.
- C3
Extension permissions, encrypted local storage, and retention/export behavior must remain understandable to the user.
Architecture
LinkedIn content scripts plus manual pasted-post input; React side-panel workflows; typed drafting/source-research ports with free OpenRouter-first routing and Gemini/Groq fallback; encrypted browser storage, migrations, backup/export, and Manifest V3 permissions.
01
Page integration
LinkedIn content scripts/extraction plus bounded manual pasted-post input.
02
Extension interface
React side-panel workflows.
03
Provider boundary
Typed drafting/source-research ports with curated free OpenRouter models, bounded provider timeouts, and Gemini/Groq fallback.
04
Storage and delivery
Encrypted browser storage, migrations, backup/export, and Manifest V3 permissions.
Decisions
Use a bounded validated untrusted-content envelope
Use a bounded, validated untrusted-content envelope to reduce prompt-injection/data-sprawl risk.
Ask for permissions when a capability needs them
Ask for permissions when a capability needs them rather than upfront, trading workflow branching for user control.
Contribution
- Implemented reply, idea, refine, history, settings, and onboarding journeys.
- Built content extraction and validated prompt envelopes.
- Added manual four-direction replies, model provenance, and visible provider recovery states.
- Added encrypted credential storage, migrations/recovery/retention, configuration backup, and data archives.
- Established static, unit, packed-extension, Axe, responsive, and visual-regression gates.
Outcomes
- Users can paste a post or select a visible conversation and draft four reply directions.
- Generate/refine post ideas in configured voice.
- Use a fixed free-only OpenRouter, Gemini, and Groq route under defined fallback rules.
- Export and restore validated configurations and history.
Product screenshot
Verified stack
- WXT
- React 19
- TypeScript
- Chrome Manifest V3
- Chrome Extensions API
- Tailwind CSS
- Hugeicons
- OpenRouter API
- Gemini API
- Groq API
- Zod
- Vitest
- Playwright
- React Testing Library
- TanStack Start
- GitHub Actions
- CI/CD
- Git
- System Design
Working through something similar?
Share your challenge, scope, and timeline to start a focused conversation.