How to implement a personalized task to Kai, skill or agent?

[lisamariet]

Say if I wanted to personalize Kai to assist me in managing my email inbox. both to catergorize and clear up mailbox but also to notify me about important emails that I must act on quickly or read asap. + tasks like create calendar events, todo items etc. Lots of different tasks. Basically want it to be an email assistant that learns as emails arrive what I want him to do with those kinds of emails. So also with the ability to learn.

How would I go about to implement something like this? Is is an Agents.. or is it a new Skill? Or possibly many skills, and 1 agent to run them all?

[chadif]

I asked PAI on how to do it and it gave me quite a lengthy answer so I'll just post the most important portions here:

📋 SUMMARY: Architectural consultation for building an email assistant skill in PAI framework

🔍 ANALYSIS: This is an excellent PAI use case - here's how to architect it following CLI-First, Skills-as-Containers, and PAI learning principles

🏗️ PAI ARCHITECTURE FOR EMAIL ASSISTANT

Phase 1: CLI-First Foundation (Weeks 1-2)

Step 1: Create Email API CLI

Use system-createcli skill

"Create a CLI for Gmail/Outlook API"

Generated CLI Structure:
${PAI_DIR}/bin/emailcli/
├── emailcli.ts
├── README.md
└── QUICKSTART.md

Commands:
emailcli inbox list --unread
emailcli read
emailcli categorize
emailcli archive
emailcli delete
emailcli search
emailcli labels list
emailcli labels apply

Step 2: Create Calendar/Tasks CLIs
${PAI_DIR}/bin/gcal/ # Google Calendar CLI
${PAI_DIR}/bin/todoist-cli/ # Or whatever task system you want to use

Following the same CLI-First pattern as llcli.

---

Phase 2: Email Management Skill (Weeks 3-4)

Create the skill:
${PAI_DIR}/skills/email-assistant/
├── SKILL.md # Main orchestration
├── workflows/
│ ├── triage-inbox.md # Process unread emails
│ ├── categorize-email.md # Smart categorization
│ ├── detect-actions.md # Find actionable items
│ ├── create-calendar.md # Email → calendar event
│ ├── create-task.md # Email → todo item
│ └── learn-pattern.md # Capture decision patterns
├── rules/
│ └── email-rules.json # Learned categorization rules
└── patterns/
└── learned-patterns.md # Decision history

SKILL.md structure:

name: email-assistant
description: |
Intelligent email management with learning capabilities.

USE WHEN user says "triage my inbox", "process emails",
"what emails need attention", "create task from email",
"learn email pattern", or requests email assistance.

---

Email Assistant Skill

Manages email inbox with categorization, action detection,
and learning from user decisions.

[chadif]

As for the design, I have mapped it out to this:

Going back to your agent question, as you can see you have to have the email skill in place to be used directly or routed to agents. In other words the email "skill" gets to decide when to route things to agents (more complex worflows), otherwise it just relies on the CLIs if it's able to accomplish the task. You can also do direct delegation from CORE to Email agents, but you'll have to modify the CORE/routing slightly to do that. I suggest you go through Email skill and then let it decide to use agents if it wants to.

[emory]

I haven't hauled up my PAI past 0.6.0 yet but my DA regularly will write code, but prefers to install binaries or executables in my user's ~/.local/ directories.

[chadif]

Not even sure what version I have lol

Currently the way it works for me:

• generate the code in .claude/bin/[CLI_NAME] using the system-createcli skill
• make CLI executable: chmod +x ${PAI_DIR}/bin/{{CLI_NAME}}/{{CLI_NAME}}.ts
• create a symlink in ~/.local/bin/ to ${PAI_DIR}/bin/{{CLI_NAME}}/{{CLI_NAME}}.ts

This way the CLI is self-contained in the .claude/bin directory yet useable from anywhere.

[smolcompute]

Why do you save your CLI tools in .claude/bin rather than in the relevant skill's folder ?

[chadif]

I didn't really come up with that design choice, the system-createcli enforces that:

from system-createcli/workflows/create-cli.md

In a way that makes a lot of sense, a CLI is a self-contained code module that can be used independently by other specialized skills, other agents, or used completely outside of PAI in terminal/bash.

In the case of email here, you can create a CLI that can read/send an email and any skill can use it through direct CLI commands as opposed to routing to an email skill for example.

[danielmiessler]

Please refer to this major change coming to the repo that should address your issues. If it doesn't please resubmit. Thank you for the work!

#222