Blogs & Webinars

Session 45: Automate Any Workflow with Agent Skills

Bala walks through Claude agent skills — markdown SOPs that turn a repeatable workflow into something Claude runs for you. Live demos cover Luma image generation, sales research, calendar integration, and audience automations.

Welcome to session 45! This week Bala Panneerselvam — Founder of Applied AI — walked us through agent skills. Skills are a deceptively simple idea: take any workflow you do repeatedly, write it down as a markdown file, and Claude runs it for you. No frameworks, no installation, no orchestrator. Just instructions.

Bala Panneerselvam is the Founder of Applied AI. He runs the AppliedAI Club community of 1,600+ practitioners and builds production AI systems for growth-stage and enterprise companies.

Session Overview

A skill is not a tool, not an agent, not a framework. It's an instructions file written in plain markdown that Claude reads when the context is right and follows when you ask. The session covered why skills exist (splitting your one giant agent.md into context-specific instructions), what goes into a skill file, how multiple skills cascade together, and three live demos pulling real work — image generation, sales research, calendar workflows — out of conversation and into automation.

Key Takeaways

  • A skill is just an instructions file: Some LLM providers have tried to take ownership of the concept and make it sound technical. It isn't. It's a .md file with a name, a description, and the steps you'd give a competent assistant. That's the whole thing.
  • Skills come from your agent.md, not your imagination: Start with the single global instruction file you (or your tool) already have. When it gets too big, split it. Each split becomes a skill. Claude picks the right one based on context.
  • Cascade skills for compound automations: One skill creates a Luma event invite image. Another skill posts it to LinkedIn. A third schedules the email. Compose them and your entire run-of-show is automated.
  • The skill builder is iteration, not architecture: Bala's typical skill takes about a month and four or five iterations to freeze. There is no upfront design. You write the first version, run it, see where it failed, fix that, repeat. After about a month, you don't go back to it.
  • Connect the tools, not the platform: Skills work across MCP-connected tools — Outlook, Gmail, Google Calendar, Linear, Jira, Nano Banana, Luma. If the tool isn't natively wired up, Claude will tell you how to connect it.
  • Keep the human in the loop for outbound writes: Bala's rule of thumb — automate reads freely, but require human approval before an LLM sends a message, posts content, or replies on your behalf.

What Are AI Agent Skills?

The reason skills exist: when you give Claude one giant agent.md with everything it should ever do, the model gets overwhelmed and starts failing. Split that one file into context-specific instructions and Claude reaches for the right one when the topic matches. That split is a skill.

A skill file has three parts:

  • Name — what the skill is called
  • Description — when Claude should reach for it
  • Tools (optional) — which tools the skill is allowed to use

That's the entire format. Everything else is the instructions you'd give a human assistant doing the same job. No special syntax, no framework, no SDK.

Live Demos

Demo 1: Luma Event Invite Generation

When you joined this session, you saw a Luma invite image — the event cover, with title, presenter photo, date. Claude generated that image automatically. The skill instruction is roughly "given the event details, create a 1080×1080 image in our brand style". Claude sends the spec to Nano Banana, gets the image back, and posts it. No manual design step.

Demo 2: Cascading Skills

A primary skill creates an image. A second skill takes that image and automatically uploads it to LinkedIn or Luma. A third triggers a notification. Cascade enough of them and your entire workflow gets automated — without a workflow engine, just a chain of plain-markdown skills.

Demo 3: Sales Research

The session's standout demo: a sales-research skill that takes a calendar meeting, looks up the prospect (Twitter, LinkedIn, founder background, public funding data), and produces a brief before the call starts. Bala walked through how to build this skill yourself — the structure of the instruction file, where to be explicit about which LLM does which step, and how to add guardrails on personal data.

How to Build Your Own Skill

The walkthrough was concrete:

  1. Write the manual version first — do the task yourself and write down every step. That's the draft instruction set.
  2. Convert to a markdown skill file — name, description, then the steps.
  3. Tell Claude which LLM to use per step — for some sub-tasks (translation, classification) a smaller cheaper model is enough. Be explicit.
  4. Run it once — see where it fails. Bala showed his Yashveer-from-the-audience example: the original automation kept double-sending WhatsApp messages until they iterated through that bug.
  5. Iterate for ~four to five rounds — typically a month of part-time use. After that, you stop touching it.

Audience Demos

Yashveer Yadav shared his daily LinkedIn-to-Excel automation — pulls activity from his network and structures it into a spreadsheet every morning. Same approach, different workflow.

He also demoed his WhatsApp birthday-wishes automation. Important detail: he keeps the human in the loop for sending. The skill drafts the message and surfaces it. He reviews. He clicks send. That's the right boundary for outbound personal writes.

Q&A Highlights

  • Skills vs agents vs MCPs: A skill is an instruction file. An agent is the LLM running with a set of skills available. An MCP is the protocol that connects external tools. You write skills, Claude is the agent, MCP brings in your tools.
  • Skill scheduling: Yes, you can schedule a skill to run on a cadence. Skill + scheduler is independent — the skill is the what, the scheduler is the when.
  • Skill architecture for a new project: There is no upfront architecture. Start with one skill, iterate it monthly, split when one skill outgrows itself. Bala explicitly avoided giving a framework — it's iteration through real use.
  • Cross-platform tool integration: If a tool isn't natively connected (e.g., Outlook calendar in a Claude session), ask Claude how to connect it. It will tell you the MCP setup.
  • Security and personal data: Read access is fine. Outbound writes (post, send, reply) keep a human in the loop. Sensitive accounts (banking, personal email) — don't connect them at all.
  • What about co-pilots and IDEs: The same skill markdown file works in Claude Code, Claude Desktop, Cursor, and most current tools. Drop it into the project, it gets picked up.

Here's the entire recording of the session.