Development plan
Message Flow Kit (powered by ai)
An inbox + reply coach that pulls messages into one place, summarizes them clearly, suggests pleasant replies, and learns how you actually talk—so people can respond without accidentally offending someone, while still sounding like themselves.
Product idea
People get messages all day across different apps. The hard part isn’t reading them—it’s answering in a way that feels kind, clear, and true to who they are. Message Flow Kit condenses incoming messages into organized, readable language, then offers a few warm response options plus a custom AI draft from the user’s own idea.
Core promise:
“Reply like you, without the stress of saying the wrong thing.”
What it does
- Brings messages into one workflowed place
- Summarizes each thread in plain, organized language
- Suggests a few pleasant replies that sound good to the other person
- Offers a custom response path: type the idea → AI drafts it in your voice
- Lets users define flows (work, friends, family, clients, etc.) and how they naturally respond in each
- Learns over time from approved replies—tone, personality, what to share and what not to share
Reply styles (v1)
- Warm / friendly
- Clear / direct but kind
- Soft decline / boundary
- Custom — user types intent (“can’t make Saturday, free Sunday”) and AI drafts in their voice
Voice & learning model
The app builds a lightweight profile of how the user communicates—not a creepy clone on day one, just something that gets closer over time.
Onboarding captures
- How they greet people
- Formality at work vs with friends
- Words they never use
- Things they never share (money, location, family details, etc.)
- How they decline, apologize, and follow up
Stored as
- Structured preferences and tone dials
- Example pairs: incoming snippet → approved reply
- “Do not share” rules
Learning loop
- Suggest replies
- User picks or edits one
- Save the approved final text as a training signal
- Prefer patterns from recent approved replies
Start with prompting + retrieval of the user’s examples. Full model fine-tuning comes later.
Flows
Flows are buckets for different social contexts. Each flow has a purpose, default tone, sharing rules, and typical response patterns.
- Clients / sales
- Friends
- Family
- Dating
- Coworkers / boss
- Support / difficult conversations
When a message arrives, classify it into a flow (AI + user correction), then draft using that flow’s rules plus the personal voice profile.
Kind by default—not fake-happy
“Always happy” can feel fake or unsafe. Better framing: kind by default, honest when needed, protective when boundaries are hit.
- Detect conflict, guilt-tripping, or money asks
- Offer warm decline templates
- Warn before oversharing
- Never invent facts (dates, promises, money)
- Ask the user to confirm anything commitment-like
Build steps
Phase 0 — Nail the product
- Ship one channel first (email, SMS, or social DMs—not everything)
- Cap v1 at three reply styles + custom intent
- No auto-send in v1 — user always approves
- Don’t claim a perfect personality clone yet; aim for “closer over time”
Phase 1 — Manual MVP
- User pastes a message or thread
- App returns: short summary, tone/risk notes, 3 suggested replies, custom-intent box
- User edits and copies the reply out
- User marks which reply they liked or sent
This proves demand before hard inbox integrations. Can live on the EVO site as an early tool while collecting waitlist emails.
Phase 2 — Identity & voice profile
- Build onboarding for tone, boundaries, and examples
- Store prefs + approved reply pairs
- Wire the learning loop from edits and selections
Phase 3 — Workflows (flows)
- Let users create and name flows
- Per flow: purpose, tone, share/don’t-share rules, response patterns
- Classify messages into flows with easy correction
Phase 4 — Unified inbox
Connect real sources after the paste MVP works.
- Email (Gmail API)
- Calendar awareness (optional, for scheduling replies)
- SMS (harder, more compliance)
- Meta DMs / LinkedIn (platform limits)
System shape:
- Ingest via OAuth / webhooks
- Normalize into one thread model
- Summarize + classify flow
- Generate reply candidates
- User approves → send or copy
- Learn from approval and edits
Privacy is first-class: encryption, retention controls, “never train on this thread,” and clear visibility into what the AI can see.
Phase 5 — Safety polish
- Boundary detection and soft declines
- Overshare warnings
- Fact-check prompts before commitments
Suggested stack
- Web app first (PWA / mobile later)
- Auth for accounts
- Neon for users, flows, threads metadata, approved examples, prefs
- Blob for attachments if needed
- AI for summarize + draft; retrieve similar past approved replies
- Background jobs for ingest and classify
- Send only after explicit user action in v1
Milestones
| Milestone | Outcome |
|---|---|
| M1 | Paste message → summary + 3 kind replies + custom intent |
| M2 | Voice profile + flows + learn from edits |
| M3 | Gmail connect + unified thread UI + approve to copy/send |
| M4 | Stronger personalization from history + boundary rules |
| M5 | Second channel + mobile polish |
Risks to plan around
- Integrations and platform policy — start with paste + email
- Trust and privacy — people won’t connect inboxes without clarity
- Creepy over-personalization — earn it with transparent controls
- Generic AI voice — learning only works if approved replies are saved
- Auto-send temptation — delay it; approval builds trust
First shippable slice
Reply Calm mini-tool: paste an awkward message → pick a flow → get 3 pleasant replies + a custom draft → save preferred style. That proves the product, feeds the waitlist, and becomes the seed of the full Message Flow Kit.