Twillot

Power Twitter/X users save thousands of bookmarks annually as a digital second brain – researchers archive studies, founders collect competitive insights, knowledge workers hoard reference threads. Yet 78% of saved content is never revisited (source: 2023 Read It Later study), creating passive archives where critical insights decay unseen. Users spend 3.1 hours/week manually searching their saves (n=142 surveyed power users, Jan 2025), costing high-value workers $8,400/year in lost productivity at blended $52/hr rates.

Business Case:
41,500 target users (source: Twillot power user cohort, Dec 2024) × 52 digest deliveries/year × $2.10 incremental LTV from reduced churn & engagement lift (source: analogous Notion AI adoption data) = $4.5M/year incremental value.
If adoption is 40% of estimate: $1.8M/year.
Cost to build: $310K (2 engineers × 18 weeks + $80K cloud/AI ops) using India regional benchmarks.

This feature is an automated weekly synthesis of personal content themes, forgotten gems, and cross-time connections. It is not real-time search, content creation, or social sharing.

Pocket surfaces recent saves but lacks thematic analysis. Readwise highlights popular annotations but ignores personal patterns. Notion AI connects documents but not tweet-level insights.

Our wedge is proactive insight generation because competitors only react to queries.

WHO / JTBD: When a founder researching competitive landscapes saves 200+ tweets/week, they want to surface latent patterns in their saved content without manual review – so they can act on insights before they become stale.

WHERE IT BREAKS: Users must scroll through chronological lists or guess search terms, missing thematic clusters and time-separated connections. Saved content lacks temporal context ("Why did I save this?"), and high-signal items get buried.

WHAT IT COSTS:

Business Impact: 41,500 users × ($8,400 + $468) = $368M/year recoverable value (source: Twillot power user cohort size, time/incident data from Jan 2025 survey).

Integration Map:

• Reads: Bookmark DB (content, save dates), Like History API, Tweet Metadata Service
• Writes: Digest Content Store (themes/gems/connections), Email Service, In-App Feed

Core Flow:

1. Nightly job processes new saves (7-day window)
2. NLP clusters content into themes using BERT embeddings
3. "Hidden gems" identified by: low revisit rate + high engagement score (likes/comments)
4. "Connect dots" finds content pairs with high semantic similarity >90 days apart
5. Digest rendered at 6am UTC every Monday

Key Decisions:

• Weekly batch (not real-time) to allow pattern accumulation
• Email-first delivery (higher open rates for async content)
• No social sharing (maintains private thinking space)

UI Wireframes:

┌──────────────────────────────────────────────┐
│ AI KNOWLEDGE DIGEST: WEEK OF APR 28          │
│ [user avatar]  •  Unread  •  Archive         │
├──────────────────────────────────────────────┤
│ 🔍 TOP 5 THEMES                              │
│ 1. AI Agent Orchestration (14 items) →       │
│ 2. VC Funding Climate (11 items) →           │
│                                              │
│ 💎 HIDDEN GEMS                               │
│ • @tomasz: "GPT-5 sparse MoE breakdown"      │
│   (saved 8 months ago, 0 views) →            │
│                                              │
│ 🧩 CONNECT THE DOTS                          │
│ "RAG evaluation frameworks" ←[6 months]→     │
│ "RAG production monitoring pitfalls" →       │
└──────────────────────────────────────────────┘

┌──────────────────────────────────────────────┐
│ EMAIL SUBJECT: Your hidden gem: VC funding   │
│                                              │
│ Hi [Name],                                   │
│ Your top theme this week: VC Funding (11)    │
│                                              │
│ 💎 You saved but never viewed:               │
│ "Seed stage dilution benchmarks" (Mar 2024)  │
│                                              │
│ 🧩 Connected: "Q1 2023 downturn" →           │
│ "Q4 2024 recovery signals"                   │
│ [View Full Digest]                           │
└──────────────────────────────────────────────┘

Phase 1 — MVP (10 weeks)
US#1 — Generate Themes

• Given 200+ saves in 7 days
• When weekly job runs
• Then system surfaces 5 themes with ≥90% accuracy against human-labeled sample (Validated by PM against n=50 user archives)
• If fails: Digest omits themes section (P1 dimension)

US#2 — Identify Hidden Gems

• Given saves older than 30 days
• When engagement score ≥ cohort 75th percentile
• Then system selects 3 gems with ≤5% false-positive rate (Validated by Data Science against n=100 gems)
• If fails: Fall back to "most engaged with old saves" (P2 dimension)

Out of Scope (Phase 1):

Phase 1.1 (4 weeks post-MVP):

• In-app digest viewer with archive
• "Save this insight" button for gems

Primary Metrics:

Guardrail Metrics:

What We Are NOT Measuring:

• Total digests sent (vanity; doesn't measure value)
• Theme count per user (quality matters over quantity)
• Email clickthrough rate (open rate is primary indicator)

Risk 1: Low Insight Accuracy

• Probability: Medium | Impact: High
• Mitigation: Data Science lead (Priya) to validate clustering model against 500 human-labeled saves by Week 6. If accuracy <85%, use hybrid human-in-loop during beta.

Risk 2: EU AI Act Compliance

• Probability: High | Impact: Blocking
• Mitigation: Legal (Markus) to confirm "limited risk" classification by Week 8. Fallback: Disable for EU users until Article 17 documentation complete.

Risk 3: Compute Cost Spikes

• Probability: Medium | Impact: Medium
• Mitigation: Infra lead (Jin) to implement per-user cost cap ($0.15/digest) by Week 9. Breach triggers digest skip for heavy users.

Kill Criteria — review if ANY met:

1. D90 open rate <25% with 40% cohort penetration
2. False theme rate >15% at D60
3. EU launch blocked at Week 10 with no mitigation path

Decision: Content processing scope
Choice Made: Bookmarks + Likes only (excludes tweets without explicit save)
Rationale: Likes are lower intent; including them would dilute signal. User testing showed 83% prefer bookmarks as "canonical saves".

Decision: "Hidden gems" definition
Choice Made: Saved >30 days ago, opened ≤1 times, engagement score ≥75th percentile
Rationale: Rejected "never opened" (too narrow) and "no engagement filter" (noise risk). Threshold balances novelty and quality.

Decision: Digest delivery format
Choice Made: Email primary, in-app secondary
Rationale: 72% of surveyed users prefer email for weekly digests (vs 28% app). In-app feed added for discovery but not as primary.

Decision: Thematic clustering depth
Choice Made: 5 themes max, 8-word max labels
Rationale: Cognitive load testing showed >5 themes reduced retention. Labels must be scannable in notification previews.

Before/After Narrative:
Before: Elena (founder) spends Thursday mornings scrolling through 300+ bookmarks to prep her investor update. She misses a critical thread about semiconductor shortages saved 4 months ago, leading to an inaccurate market analysis. Her engineering lead later finds it, costing 8 hours in rework.

After: Elena’s Monday digest surfaces "semiconductor shortages" as a top theme with the forgotten thread flagged as a hidden gem. The "connect dots" section links it to a recent save about factory reopenings. She integrates both into her update in 9 minutes, impressing investors with forward-looking analysis.

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

1. Users perceived themes as inaccurate because we didn't tune for sarcasm/niche jargon, leading to 32% unsubscribe rates by D60.
2. The 'connect dots' algorithm failed with short-form content, producing trivial connections that users dismissed as noise.
3. Compute costs hit $0.38/digest for power users, making the unit economics unsustainable at target adoption.

Success looks like: A research director tweets "Twillot Digest found the paper that got me funded". Support tickets for "lost saves" drop 65%. The CFO notes at QBR: "This moved our power user NPS from 31 to 44 – double the retention budget impact we modeled"."