Daily Pulse Internals

This page explains why Daily Pulse can proactively surface "what's worth looking at today" and which real signals it depends on.

Positioning: not a news feed, not another chat tab — a proactive curation layer.

Read First

We recommend trying Daily Pulse first (generate manually a few times, give some feedback). Concepts like "feedback profile", "open tasks", "tomorrow curation" stay abstract without hands-on.

The Real Problem Daily Pulse Solves

Most AI products assume the user knows what to ask. The reality often is:

• The user just opened the app and hasn't formed a question
• The user knows they're busy but doesn't know which thing to push next
• The user's long-term interests, tool capabilities, and recent chats are enough to suggest a few high-value directions

Daily Pulse front-loads that step.

It's Not a "Cloud Recommendation System"

Daily Pulse's signal sources are local context and locally-already-fetched results, not a server-side profile.

It doesn't rely on a unified account profile for cross-user recommendation. It curates around this device's actual usage trail. This means:

• It's closer to a personal curator, not a public feed
• Its quality depends heavily on your real usage and feedback discipline
• It can't know external facts out of nothing — only if you gave context, or recently fetched MCP / Shortcut output

The Eight Input Sources

A DailyPulseGenerationInput is built at generation time, with eight categories.

1. User-Authored Focus

If you wrote "what to focus on today" in Daily Pulse, it becomes a top-priority personal signal.

2. "What I Want Tomorrow" Curation

Not an instantly-effective note — it's curation aimed at the next batch.

Good for:

• Direction to keep advancing tomorrow
• Topics to dial down
• Card types to dial up

3. Recent Chat Excerpts

Daily Pulse doesn't pour all history in. It takes truncated slices of recent sessions.

Current defaults:

• Up to 4 sessions
• Up to 6 user/assistant messages per session

The goal isn't full recall — it's understanding what you've been busy with.

4. Long-Term Memory

It reads un-archived long-term memories, capped.

Current default: at most 8 entries, length-compressed.

This lets Daily Pulse remember:

• The content types you favor over time
• The projects you keep pushing
• Topics you consistently care about

5. Recent Request Log Summary

The last 7 days of activity get squeezed into a short summary:

• Total request count
• Frequently-used providers
• Frequently-used models

Not a tech stats report for the model — just letting it know what you've been leaning on.

6. Open Pulse Tasks

A card you convert into a task doesn't stay a one-off recommendation — it participates in later generations.

Design intent:

• Daily Pulse covers discovery and follow-through
• But doesn't robotically repeat the old card title — it leans toward "help push one step forward"

7. Feedback Preference Profile

Historical feedback is shaped into three prompts:

• Topics you're more likely to want
• Topics to avoid
• Topics that already showed up in recent days

Sources: manual likes / downvotes / hides, plus "save as conversation" as a strong positive signal.

8. External Capabilities and Results

The most often-misunderstood layer. Daily Pulse distinguishes three things:

Type	Meaning
Available capabilities	Which MCP servers are selected for chat, which Shortcuts can be called
Recently-fetched external result snapshots	Most recent Shortcut result, most recent MCP output
Announcements & trend signals	App announcements, trend-style signals

A Hard Constraint

Tool capability descriptions only mean "can be called" — they do not mean live data has been fetched.

Only "recently-fetched external result snapshots" represent external content the user actually has.

Defaults and Retention

Concrete defaults in the current implementation:

Item	Default
Auto-generate	On
MCP context	On
Shortcut context	On
Recent external result snapshots	Off
Trend / announcement signals	On
Final card count	3
Model candidate count	6
Run retention	14 days
Feedback history retention	120 entries
External signal retention	40 entries
Task retention	80 entries

Also note:

• Persistence keeps multiple recent days
• The UI shows only "today" by default

Historical batches feed feedback and preference accumulation, not a timeline feed.

Model Selection Priority

Daily Pulse tries the dedicated model first:

1. The user's explicitly-set Daily Pulse Model
2. The current chat model, if it supports chat
3. The first chat-capable active model

This lets you tune Daily Pulse independently without it silently breaking when no dedicated model is set.

Prompt Strategy

The Daily Pulse system prompt doesn't free-roam. It spells out the job and boundaries.

Core rules:

• Curate based on recent chats, long-term memory, usage trail, feedback, external capabilities, and user-authored input
• Prefer pushing open tasks forward, don't restate old titles verbatim
• Cards must be specific, continuable, and savable as a new session
• Don't fabricate user history, don't invent external facts
• With MCP / Shortcut capabilities, prefer cards that can be immediately advanced
• Tool capability ≠ live data fetched
• Output must have diversity — three cards shouldn't all orbit the same thing
• Topics the user has clearly disliked → minimize

Output Schema

The model must return JSON only, no Markdown code fences.

Fixed structure:

{
  "headline": "single-sentence top headline",
  "cards": [
    {
      "title": "card title",
      "why": "why this is for the user",
      "summary": "one or two sentences",
      "details_markdown": "detailed Markdown that can be saved as the first chat message",
      "suggested_prompt": "if continued, what to send"
    }
  ]
}

Five fields solve five problems:

Field	Role
title	What the user sees first in the list
why	Why you specifically, not random
summary	Decide at a glance if it's worth opening
details_markdown	First-message content if saved as a session
suggested_prompt	Quick follow-up without re-thinking the question

Model Proposes, Client Filters

Daily Pulse doesn't trust the model's raw output verbatim. Actual flow:

1. Ask the model for 6 candidate cards
2. Client does normalization, length trimming, fallback
3. Client scores by preference profile and filters down to 3

Local Filtering Rules (Explainable Scoring)

Rule	Delta
Near user's current focus	+4
Hit past positive feedback topics	+3
Hit topics shown recently	-2
Hit past negative feedback topics	-6
Details are substantial	+1

Then three more steps:

• Drop cards too similar to negative-feedback topics
• Drop cards too similar to each other
• Avoid all three falling into one category

That's why Daily Pulse looks like "model-generated" but carries a clear product-curation flavor.

Feedback Loop

A card can receive four classic feedback types:

• Liked
• Downvoted
• Hidden
• Saved as conversation

These don't train a remote recommender. They directly become prompt input for the next local curation.

Daily Pulse's "learning" is more next-round prompt hinting, not an opaque weight update.

Why "Save as Conversation"

A key design.

Saving a card creates:

• A new session named after the card title
• The card's details_markdown as the first assistant message
• A record of this action as strong positive feedback

So Daily Pulse isn't a terminal — it's a chat-entry generator.

Why "Convert to Task"

Some cards don't need chat right now; they need follow-up.

Tasks preserve:

• Title
• Summary
• Suggested prompt
• Completion state

Later generations factor open tasks into curation.

Why Delivery Isn't Guaranteed

Daily Pulse alerts and delivery are local capabilities, not server push.

Current defaults:

• Morning reminder off by default
• Default reminder time 08:30
• iPhone prepares today's batch ahead of time when possible
• If reminder time hits before prep finishes, send a "ready" notification
• Apple Watch leans on foreground-resume to start prep

The whole design is best-effort:

• Respect iOS scheduling boundaries
• Don't pretend to be an always-on cloud cron

How to Accurately Think About It

The most accurate framing isn't "daily recommendations" but:

A proactive curator driven by recent chats, long-term memory, usage trail, external capabilities, feedback history, and open tasks.

Its value isn't "think for you" — it's reorganize context you've already accumulated but haven't structured into a few actionable entry points.

Next

• See how memory / summary / profile each work → Memory, Summary & Profile
• See worldbook + tool governance details → Worldbook & Tool Governance