Example Packs

A pack is judged by whether it keeps a real user coming back for weeks. The way to evaluate the format is to look at three concrete shapes a serious vertical agent app might take.

This page sketches each one — its memory, its tables, its cron jobs, and the skills that move data through it.

English Coach

“Remember what I got wrong last week and quiz me on it this morning.”

A pack that runs a long-term English-learning loop. The user practices daily; the pack tracks what they confuse and surfaces it in tomorrow’s mission.

What it remembers

memory/
├── soul.md     — Coaching tone: encouraging, specific, not patronizing
├── user.md     — Stable: CEFR level, target (IELTS), available minutes/day
└── state.md    — Evolving: this week's focus, recent mistake patterns

schema.sql
├── writing_submissions   — every paragraph the user has written
├── review_queue          — SRS items with phrase, due_at, ease_factor
├── mistake_patterns      — recurring error types (article use, tense agreement)
└── weekly_progress       — week-keyed summary rows

Skills

Skill	Trigger	Writes
onboarding	First open	user.md (CEFR, goals), seed state.md
daily_mission	9am cron, manual	writing_submissions
writing_feedback	After submission	review_queue, mistake_patterns
vocab_review	Manual, button	Updates review_queue.ease_factor
weekly_report	Sunday 8pm cron	weekly_progress
aggregator	Weekly	Updates state.md summary

Cron

cron:
  - { id: daily_practice, schedule: "0 9 * * *", skill: daily_mission }
  - { id: weekly_review,  schedule: "0 20 * * 0", skill: weekly_report }

Why it works

Everything the agent says next morning is grounded in review_queue and mistake_patterns, not the conversation history. The user can open state.md and read exactly what the pack thinks of them. If something is wrong, they can edit it.

Creator Studio

“Capture this link, draft three angles for it, slot the best one on Thursday.”

A pack that turns raw inputs (URLs, ideas, screenshots) into a publishing pipeline. The user feeds it material; the pack drafts, scores, and schedules.

What it remembers

memory/
├── soul.md     — Writing voice: tight, opinionated, no fluff
├── user.md     — Stable: target platforms, audience, tone constraints
└── state.md    — Evolving: this week's themes, posts in flight

schema.sql
├── sources         — captured URLs, notes, attachments
├── ideas           — angles extracted from sources
├── drafts          — drafts with status (queued / scheduled / published)
├── publishing_log  — what shipped where, when
└── brand_voice     — accepted/rejected drafts → tone signal over time

Skills

Skill	Trigger	Writes
capture	Inbound link or note	sources
ingest_url	Async after capture	Enriches sources with summary
prompt_me	Manual, button	ideas from existing sources
draft_post	Manual, button	drafts with platform variant
score_draft	After accept/reject	Updates brand_voice
weekly_plan	Sunday cron	Reorders the drafts queue

Cron

cron:
  - { id: morning_idea,   schedule: "0 8 * * 1-5", skill: prompt_me }
  - { id: weekly_plan,    schedule: "0 18 * * 0",  skill: weekly_plan }

Why it works

The pack does not “be a writing assistant.” It runs the loop the user actually does: capture → angle → draft → ship → review. Each step writes structured data that the next step reads. A general chat assistant has to be re-told all of that every conversation.

Daily Journal

“Two weeks of notes — what kept coming up?”

A pack for reflective journaling, mood tracking, and weekly review. Lighter on tables than the other two, heavier on memory.

What it remembers

memory/
├── soul.md     — Companion tone: warm, curious, not therapeutic
├── user.md     — Stable: rough rhythm, life areas being worked on
└── state.md    — Evolving: themes from the past 2-4 weeks

schema.sql
├── daily_notes     — date, content, mood tag, energy
├── goals           — long-running goals with status
├── action_items    — extracted next-actions from notes
└── weekly_reviews  — synthesized weekly summary rows

Skills

Skill	Trigger	Writes
daily_note	9pm cron, manual	daily_notes
extract_actions	After note	action_items
review_week	Friday evening cron	weekly_reviews
theme_finder	Manual	Updates state.md themes

Cron

cron:
  - { id: evening_prompt, schedule: "0 21 * * *", skill: daily_note }
  - { id: friday_review,  schedule: "0 19 * * 5", skill: review_week }

Why it works

The user does not have to “remember to journal.” The pack opens the conversation at 9pm. The user does not have to “write a weekly summary.” The pack offers one on Friday based on the past 7 days of daily_notes. The memory contract lets the agent stay coherent across weeks.

What unifies the three

Property	Why all three need it
Typed state	The thing the user keeps coming back to is the accumulated state, not the last reply
Per-pack memory	soul.md/user.md/state.md keep each pack in its own voice
Schedule	The pack opens the conversation; the user does not have to remember to engage
Permission scope	Each pack only touches what its pack.yml declares
Eval surface	Each pack can be regression-checked: did the SRS scheduler still pick the right cards? did drafts still match the brand voice?

If a vertical agent doesn’t need at least four of those, it probably doesn’t need to be a pack. A one-shot tool does not need a .aouo bundle.

Building your own

These three are sketches — they are not shipped as code in this repo. The samples actually in apps/ (notes, create, vocab) are earlier and smaller versions of the same idea.

To start: see Your First Pack, then Pack Spec and Schema & Persist.

Related

• Why Packs — what the pack boundary buys
• Five Pillars — what a serious pack should implement
• Builder Direction — the long-term goal of NL → pack