ADR-0001: AI Assistance Philosophy
Status: Accepted
Date: 2026-01-05
Context: Memory Maker helps people write tributes for loved ones. The core tension: most people struggle to articulate meaningful sentiments, but the output must feel authentic to the contributor—not AI-generated. Sheri explicitly stated she doesn't want tributes "written by AI."
Decision Drivers:
- Tributes must feel authentic to the person contributing
- Most people are poor at expressing sentiments in writing (why they don't write letters/notes)
- Competitors don't offer AI-assisted writing
- Users are savvy enough to detect AI-generated content
Decision
AI assistance will be assistive, not replacement—similar to Grammarly rather than ChatGPT.
The AI will:
- Prompt users with contextual questions based on relationship and occasion
- Offer light cleanup suggestions (grammar, clarity, tightening)
- Provide gentle guidance without rewriting
The AI will NOT:
- Generate tribute content
- Rewrite user submissions
- Replace the contributor's voice
Rationale
Sheri's exact framing: "I love the idea of prompting and helping people get there and formulate. I don't want it to be written by AI."
The value proposition is helping people do something hard—expressing love and appreciation in writing—not doing it for them. The emotional weight of a tribute comes from knowing a real person wrote it. AI-generated content would undermine the core product value.
The Grammarly model provides the right mental framework: enhance what the user wrote, don't replace it.
Considered Options
Option 1: Full AI Generation
User provides a few facts, AI writes the tribute.
Pros:
- Easiest for users
- Consistently polished output
Cons:
- Destroys authenticity (the entire value proposition)
- Recipients would feel cheated
- Sheri explicitly rejected this
Option 2: Assistive AI (Grammarly-style)
AI prompts, suggests, and polishes—user writes.
Pros:
- Preserves authentic voice
- Helps struggling writers
- Differentiator vs. competitors
- Aligns with Sheri's vision
Cons:
- More effort for users
- Output quality varies by contributor skill
Option 3: No AI at All
Pure manual entry with no assistance.
Pros:
- Maximum authenticity
- Simplest to build
Cons:
- Users get stuck staring at blank page
- Quality of output varies wildly
- "You're great! Love you!" problem
Consequences
Positive
- Product feels authentic and emotionally meaningful
- Clear differentiation from competitors
- Builds trust with users who are wary of AI
- Positions well for the target market (milestone celebrations, memorials)
Negative
- Some users may still struggle even with prompts
- More complex to implement well (prompts must be good)
- Mitigation: Iterative prompt refinement based on beta tester feedback
Neutral
- Marketing must clearly communicate "AI-assisted" vs "AI-generated"
Implementation
Phase 1 (MVP):
- Contextual question prompts based on relationship type and occasion
- Basic cleanup suggestions (optional, user can decline)
Phase 2:
- Grammarly-style inline suggestions
- Dial for assistance level (user preference)
Files Affected:
- Contributor form component
- AI prompt engineering
- Cleanup API endpoint
Related Decisions
- ADR-0004: Revision Workflow (admin editing constraints)
References
- 20260105-napkin-session.md - Original discussion
- Grammarly product model
Author: Bert Carroll Reviewers: Sheri Dudley Last Updated: 2026-01-05