Pimp

Brains are not enough.

Shadi Sharaf

Astronaut

Frontbar - Christmas edition

Level up

your AI

👷🏻‍♂️Senior Engineer, 20+ years in Web dev, 15'ish in WP
🗣️ Speak about Standards, Debugging, Performance, Testing, AI
✈️ Travel junkie
🍝 Foodie 🍣🍤🥗
🧑🏻‍💻 Agencies: XWP, Human Made, The Times.
📦 Projects: WP Coding Standards, Stream plugin, and others.
🤝 Clients: Enterprise publishers, Universities, Banks, Tech

Who am I?

https://slides.com/shadisharaf

Learn the lang
Hire our team
Give them tools
Teach them the ways
Keep them updated
Onboard them to our project

Agenda

Learn the lang

Step 0

Introduction

aka

An agent is an LLM model, who can do more than just talk, it can use tools to retrieve information and execute actions.

Agent

MCP (model context protocol) is the protocol for tool calling, MCP servers are tools exposed to Agents to use. Tools are simple functions.

Agent: "I need to request JIRA issue number: ZD-101"

Client/IDE: "Sure. [calls jira_mcp.fetch_jira_ticket(ZD-101)]

MCP: "jira_mcp.fetch_ticket = fn(ref) => fetch(jira.com/issues/zd-101)"

Client/IDE: "Here you go: [result]"

MCP

Context is what information the agent has at once point of time, aka the conversation.

Context is everything you have in your open chat, plus any rules the agent has at any point of time.

Good context = good output

Too much context = Attention loss / Hallucination

Context

Set of reusable / composable instructions to carry out

a series of actions autonomously.

Workflows / Commands

/jira

- Fetch the reference JIRA ticket and all its dependencies

- Cache it in .ai/jira/{ticket}-{date}.md

/jira-plan

- Ask busboy agent to `/jira {ticket}`

- Ask architect agent to `plan {ticket}`

An MCP server to keep.. memories. Save, update, and recall them as needed, based on context.

OR a simple folder on your project, aka Documentation.

Memory

You: Next time, do not add priority 10 to WordPress hook registration unless you need to change the number of arguments passed to the function.

Agent: Sure. {tool_call create_memory("...")}

Client: "memory.create_memory("...")"

MCP: "create_memory = fn(memory) => backend.add("..")"

Agent: Memory updated.

Hire your team

Step 1

Agents

aka

If you're assigned a new project, and tasked with hiring your own team, who'd you hire?

Architect

- Thinker

- Asks questions

- Detail oriented

- Writes comprehensive plans

Researcher

- Searches online for solutions to a problem

- Explores documentations and resources

- Compiles research reports and recommendations

Coder

- Sticks to standards

- Sticks to plans

- Updates documentation, writes inline docs

- Commits to DRY, YANI, KISS techniques

- Writes small, testable functions

QA engineer

- Uses Gherkin syntax to write QA steps

- Uses playwright / browser to run QA steps

- Attentive to details

- Thinks about edge cases

- Does not update code

- Happy with writing failing tests to pass it back to engineers

Code reviewer

- Dives into the code, examine relationships, traces code execution paths

- Uncovers possible bugs, unhandled exceptions, probable edge cases

- Compiles comprehensive reports, with recommendations

Why so many agents? Can't i just use one?

Short answer: Context, and speciality.

Give them tools

Step 2

MCPs

aka

What tools do you use that can be helpful to engineers to plan and code?

File system

Reads and writes to file system

Terminal

Uses the terminal to execute commands.

Memory

Semantic and lexical memory store

JIRA

Reads JIRA tickets, comments, meta information.

Updates tickets as requested.

Context7

Fetches updated documentation for any library

(from a dedicated database)

Perplexity

Researches topics online and provides detailed reports.

Playwright

Uses playwright to control your browser and browse through pages

Clicking and typing to get a job done / test workflows.

Subagents

Spins separate agents with separate contexts to run specific actions.

Notebook LM

Asks Google NotebookLM questions, returns with their answer.

Productive.io

Reads / creates time entries.

Figma

Reads Figma designs.

Can i add as much tools as I want? 😮

Short answer: No. There is a limit to how many you can expose at one time, before the agent ignores them.

Also: Context.

Teach them the ways

Step 3

Workflows

aka

What workflows do you find yourself doing over and over again?

Plan

Read JIRA tickets, comments, meta information.

Plan user stories, system design, and implementation documents.

Commit

Commits using conventional commits

Writes detailed PR description based on changes.

Pushes the branch.

Create the PR using PR template.

Address feedback

Checks GitHub PR feedback.

Plans fixing plans for high-certainty tasks.

Asks questions about unclear tasks.

Dev notes

Writes clear dev notes based on ticket AC.

Reads code and understands patterns.

QA notes

Writes clear QA notes based on ticket AC / Dev notes.

Writes in Gherkin / EARS format.

What to use for what

Optimizing

aka

Agents.md

Global behavior for ALL agents and workflows

Non-domain specific

Project-context

Agents

Generic behavior patterns (not a workflow)

Specific domain (Frontend/Backend/Research)

Specific model (think speed and cost)

Commands

Workflows, set of chained actions

Repetitive tasks

Remember, it's all about rules in text files. Optimize them for better context & quality, and faster workflows.

Ask yourself:

- Global behavior / all agents / project context: AGENTS.md

- Domain-specific rule: Agent

- Workflow / set of actions: Command

- Project/Feature/Folder specific rule: nested AGENTS.md

Tips

Okay, what's next?

Experiment: Try, track, train, iterate.
Onboard: Prime your agents with project context, in AGENTS.md
Document: Keep comprehensive docs, saves time / enhances quality.
Memory: Keep planning docs, cache JIRA tickets, keep track
Globalize: Globalize non-project rules/agents/commands.
Localise: Share with the team, embed into repos.
Models: Pick the right models for agents (dump, thinking, etc). **
Share: Learn in the open, share what works, ask questions.
Plan: ALWAYS plan before execution, take the time.
Test: Write tests, they're cheaper now!
Review: ALWAYS review the code.

Final tips

Spend the effort!

Couple of hours today

saves

Couple of hours everyday

Share your team and prompts in #talk-ai

🧵 Agent: Coder

🧵Command: Planner

🧵Model review: Composer

Pimp

Level up

your AI

Who am I?

Agenda

Learn the lang

Introduction

Agent

MCP

Context

Workflows / Commands

Memory

Hire your team

Agents

If you're assigned a new project, and tasked with hiring your own team, who'd you hire?

Architect

Researcher

Coder

QA engineer

Code reviewer

Why so many agents? Can't i just use one?

Give them tools

MCPs

What tools do you use that can be helpful to engineers to plan and code?

File system

Terminal

Memory

JIRA

Context7

Perplexity

Playwright

Subagents

Notebook LM

Productive.io

Figma

Can i add as much tools as I want? 😮

Teach them the ways

Workflows

What workflows do you find yourself doing over and over again?

Plan

Commit

Address feedback

Dev notes

QA notes

What to use for what

Optimizing

Agents.md

Agents

Commands

Remember, it's all about rules in text files. Optimize them for better context & quality, and faster workflows.

Ask yourself:

Action time! 🎉

Demo

Tips

Final tips

Spend the effort!

Share your team and prompts in #talk-ai

Thank You!

Questions?

Shadi Sharaf

/shadyvb

/shadyvb