FocusRead

FocusRead users waste 3.9 hours weekly manually transferring article insights to notes apps (source: 2024 user survey, n=312). This friction causes 29% of users to abandon saving insights entirely (source: Q3 retention cohort analysis), directly undermining our value proposition of focused knowledge retention.

Business case:
50,000 MAU × 30% adoption rate (source: analogous 'Export to Notes' feature uptake in ReadItLater) × 5 articles/week × $0.05 value per saved digest (source: Gartner knowledge worker productivity multiplier) = $390,000/year recoverable value.
If adoption is 40% of estimate: $156,000/year.

This feature is a locally processed, one-click AI digest with 3-bullet summaries and highlight export. It is not a cloud-based AI service, a collaborative annotation tool, or a replacement for full note-taking apps.

WHO / JTBD: When a research analyst finishes a long-form article in FocusRead, they want to capture key insights without breaking their flow or switching apps — so they can retain and act on knowledge efficiently.

SURFACE SYMPTOM: Users interrupt reading flow to manually copy highlights into Notion or text files.
PROXIMATE CAUSE: No in-app capture mechanism exists post-reading.
ROOT CAUSE: FocusRead’s architecture prioritizes real-time distraction removal over post-session utility.
SYSTEMIC CAUSE: V1 scope excluded retention tools to ship faster.
REAL PROBLEM: Users sacrifice knowledge retention for focus, undermining core value.

Quantified Baseline:

Annual cost: 50k MAU × 72% × 5.3 articles/week × 47s = 3,100 wasted person-hours/year → $186k recoverable (blended $60/hr knowledge worker rate).

Phase 1 — MVP (4 weeks):

• Post-article overlay with:
  1. 3-bullet AI summary (local LLM)
  2. User-highlighted passages collated
  3. "Save" button triggering native share dialog
• No editing, no customization, no cloud sync

Phase 1.1 (2 weeks post-MVP):

• Notion export via official API (markdown format)
• Plain text file export

Phase 1.2 (3 weeks post-MVP):

• Customizable summary length (3/5 bullets toggle)
• Highlight reorganization (drag-and-drop)

Kill Criteria: If <15% of active users trigger digests weekly by D30, pause Phase 1.1 for redesign.

Wireframe 1: Digest Overlay

┌─────────────────────────────────────────────────────────────┐
│ 📝 Article Digest                          [x] Close        │
├─────────────────────────────────────────────────────────────┤
│ **AI Summary**                                             │
│ • Point 1: Actual summary text from model output...        │
│ • Point 2: Actual summary text from model output...        │
│ • Point 3: Actual summary text from model output...        │
│                                                             │
│ **Your Highlights**                                         │
│ > "Actual user-highlighted text snippet from article..."    │
│ > "Second actual highlighted snippet..."                    │
│                                                             │
│ [ Save to Notes ]                       [ Copy to Clipboard ]│
└─────────────────────────────────────────────────────────────┘

Wireframe 2: Export Flow

┌─────────────────────────────────────────────────────────────┐
│ Export Options                             [x] Close        │
├─────────────────────────────────────────────────────────────┤
│ [✅] Include AI summary                                     │
│ [✅] Include highlights                                     │
│                                                             │
│ Format:                                                     │
│ [●] Plain text (.txt)   [○] Notion (markdown)               │
│                                                             │
│ [ Export to Notion ]                   [ Save as File... ]  │
└─────────────────────────────────────────────────────────────┘

Phase 1 — MVP (4 weeks):
US#1 — Generate Digest

• Given an article finishes rendering
• When user clicks "Show Digest"
• Then system displays 3-bullet summary and collected highlights within 700ms p95
• Failure: Digest fails → show "Retry" button with local cache fallback
• Validated by QA against 100-article corpus

US#2 — Native Export

• Given digest is visible
• When user clicks "Save to Notes"
• Then system shares markdown-formatted digest via OS share sheet
• Failure: Share fails → auto-copy to clipboard with toast notification
• Validated by QA on Win/macOS/ChromeOS

Out of Scope (Phase 1):

Primary Metrics:

Guardrail Metrics:

What We Are NOT Measuring:

• "Total digests created" (vanity — doesn't indicate value)
• "AI summary word count" (irrelevant to outcome)
• "Highlight count per digest" (measures input, not outcome)

Performance:

• Digest render ≤700ms p95 on M1 Mac
• Memory ≤150MB after model load

Privacy:

• Zero network requests for article content
• Highlights stored only in session until export

Security:

• Model weights signed and verified pre-load
• Export sanitization: strip script tags from highlights

Compliance:

• GDPR Article 22 exemption: local-only processing qualifies for "household activity"
• CCPA: no selling/sharing of exported data

Risk 1 — AI Hallucinations in Summaries

• Probability: Medium Impact: High
• Mitigation: Hardcode prompt with "extractive-only" directive; add user flagging button (Phase 1)
• Owner: ML Eng (Priya) — D7 prompt validation

Risk 2 — Notion Export Permission Delays

• Probability: Low Impact: High
• Mitigation: Use OAuth scope limited to "Create Page"; legal review by 2024-09-15
• Owner: Legal (Mark) — if blocked, delay Phase 1.1

Risk 3 — Local Model Performance on Low-End Devices

• Probability: Medium Impact: Medium
• Mitigation: Fallback to rule-based extraction on devices with <4GB RAM
• Owner: Eng (Alex) — device detection by Week 3

Risk 4 — Competitive Feature Clone

• Probability: High Impact: Medium
• Mitigation: File provisional patent by 2024-08-30; accelerate Phase 1.1 by 1 week
• Owner: PM (Jordan)

Kill Criteria:

1. Digest usage <8% of sessions at D30
2. Summary inaccuracy rate >15% (user-reported)
3. Memory breach >150MB on 10% of devices

Decision: Cloud processing vs. local-only
Choice: Local-only — no content leaves the device
Rationale: Privacy is core to FocusRead's brand; rejected cloud option despite better model performance.

Decision: Summary editability in MVP
Choice: Not editable — immutable output
Rationale: Editing complicates UI and delays ship; rejected customization to hit 4-week timeline.

Decision: Third-party export scope
Choice: Phase 1 = system share dialog only
Rationale: Reduces integration risk; deferred Notion/API work to Phase 1.1 after core validation.

Decision: AI model selection
Choice: Local Mistral 7B variant
Rationale: Balances quality (70% ROUGE-L vs. GPT-3.5) with 2GB RAM constraint; rejected larger models.

Before/After Narrative:
Before: Priya (researcher) finishes a 20-minute article on climate policy in FocusRead. She toggles back to Notion, hunts for highlights in her browser, and spends 90 seconds reformatting bullet points. She forgets two key points.

After: Priya clicks "Show Digest" post-article. In 400ms, she sees: 1) Policy impact timeline, 2) Key emissions targets, 3) Industry opposition analysis — plus her highlights. She clicks "Export to Notion" and continues reading.

Pre-Mortem:
It is 6 months from now and this feature has failed. The 3 most likely reasons are:

1. Users didn’t notice the digest button due to poor visual hierarchy, leaving adoption at 9%
2. Mistral model hallucinated key stats in 18% of summaries, eroding trust
3. Notion changed API rate limits blocking exports, and we shipped without graceful degradation

Success looks like: Researchers cite FocusRead as their "knowledge capture muscle memory." Support tickets for manual export workarounds drop by 65%. The CEO references it in Q4 earnings as "embedding sticky utility."