Dhwani AI Playground

Watch the Session

Watch the Full Session Recording

Dhwani RIS members only · ~60 minutes

Session 5 opening — the S5 page on screen, 20+ attendees in gallery view
The room, live. 20+ teammates in the MEVN room and remote, the S5 page on screen, Nihaan opening with the week’s recap and today’s agenda.

What Actually Happened

Five sessions in, and this one was the proof. Ankit walked the room through four concepts that run the AI stack — Skills, MCP, Plugins, Agents — in roughly seven minutes. Then three builders showed what they’d made with them. No slides of theory. No highlight reels. Real client work, real screens, real answers to the tough questions. Plus one live teaching moment from Abhinav on design as a skill — which turned out to be the rule of the hour.

You’ll walk away with

  • A plain-language mental model of Skill vs. MCP vs. Plugin vs. Agent — and when to reach for each
  • The four skill tiers (Foundation → Partner → Organisation → Team) and why Dhwani needs to build its own layer now
  • Ravi’s timesheet + project-governance MCP — Microsoft 365 activity auto-posted into Frappe HRMS with no custom code
  • Kashish’s ATMA Education Foundation build — CLI-zero to a full Frappe project with verticals, cohorts, web forms, and document review
  • Arpita’s Welspun Help Center — org-level resources + role-gated FAQs inside a live mGrant deployment, installable in five minutes
  • A teaching moment on why the mGrant design system skill belongs in every project MD by default

Opening — Why This Session

Nihaan opened with a candid recap. The first four sessions were about lowering the barrier — what is vibe coding, why GitHub matters, how to set up CLI, hooks, and security. The pattern he’d noticed: people were solving real problems in isolation. The point of this program is to drag those wins into the open.

This week’s shift: from “AI use” to “best AI use”. Most of the company uses Claude or ChatGPT through the UI. Far fewer use the CLI. And within the CLI, the difference between an average operator and a great one is the four concepts covered today.

“This is not a talk, it is a discussion. Stop us. Ask. Engage.” — Nihaan, opening the session

The Four Concepts Ankit’s Explainer

Ankit’s framing, in plain language: AI is a smart assistant. The four concepts are four ways you extend that assistant. They are not alternatives — they are layers. Each one does one job, and together they turn a generic chatbot into an org-aware teammate.

Four Concepts framework: Skill, MCP, Plugin, Agent — shown on screen during Ankit’s explainer
Four Concepts, One Operating System. On screen during Ankit’s explainer. Each card lists what the concept is, what it does, and how it differs from the others.
📄
Skill
A prepared knowledge base. AI’s “how to” for a specific task — written once, used every time the trigger matches.
Ankit’s example: “Pooja’s email-signature skill.” When Pooja writes a mail, the skill makes sure her signature and tone are right without her having to retype it.
Knowledge Procedural Reusable
🔒
MCP
An action connector. Gives AI live access to a system — mail, calendar, HRMS, a database. The skill knows how; the MCP gives the reach.
Ankit’s example: “A skill tells AI how to write a mail. MCP is what connects it to your inbox so it can actually send it.”
Access External systems Live data
📦
Plugin
A bundle of MCPs and skills, packaged so an admin can install it once and the whole team gets the capability. Removes the manual setup.
Ankit’s example: “Instead of everyone wiring up Office 365 MCP + the email-signature skill themselves, one plugin does it.”
Bundled Admin-deployed Turnkey
🤖
Agent
A smart worker that picks the right skills, uses the right MCPs, and runs the task end-to-end. Not a chat — a delegation.
Ankit’s example: “Give the agent a task: send a mail. It picks the email-signature skill, the MCP for your inbox, writes it in Dhwani brand voice, and sends.”
Autonomous Orchestrates Delegation model

Nihaan’s one-liner to close the segment: “MCP gives access to your system. Skills give AI knowledge. You need both. Neither is better — they are complementary.”

The Four Skill Tiers — and Why Dhwani Needs Its Own

Ankit laid out three types of skills in the call; the fourth layer is the one that ties the stack together. A skill you stack at the top inherits everything below it. The more tiers active, the more context-aware your AI becomes on day one.

Foundation Partner Organisation Team
Team Skills
e.g. “how we write release notes on our team”
Most Specific
Your Team
Organisation Skills
e.g. mGrant coding hygiene, design system, CSR testing standards
Dhwani RIS
All teams
Partner Skills
e.g. Microsoft, Superpowers, Notion partner skills
Ecosystem
Any org
Foundation Skills
e.g. frontend-design — built by Anthropic
Anthropic
Universal
Nihaan’s ask to the room

“Today as an organisation we don’t have Dhwani-level engineering skills. We need them. Coding practices. Testing practices. The ask from this session is: let’s go back and build common skills, in one common place — GitHub — so Deepak’s coding skill, Anup’s coding skill, are all available to the rest of us.”

The invitation is explicit: if you’ve solved something with a skill, contribute it. The organisation tier of this pyramid is ours to fill.

Anatomy of a Skill

Ankit walked through what actually goes in a skill file. This matches the format Anthropic documents and what the Superpowers partner skills use.

01
Name & description — short label plus a precise trigger. “Use when drafting a client release-note email.” This one line decides when the skill activates.
02
Trigger condition — the signal that the skill applies to the current task. Keep it specific; generic triggers cause noisy activations.
03
Context — why this skill exists and the domain knowledge behind it. AI needs the purpose, not just the procedure.
04
Procedure — step by step, with decision points explicit. “If X, do Y. If Z, ask first.”
05
Examples — at least one real sample of the correct output. AI calibrates quality against examples, not descriptions.

Ankit’s honest note: “Skill banana is not a one-shot. You iterate. You train it. The first version will be wrong — that’s the point.”

Live Builds Three Spotlights

AI + skills + MCPs, applied to real client work already shipping. Watch for how each builder used different parts of the stack.

01

Ravi Jangir Flagship MCP Demo

Timesheet automation & project governance — M365 → Claude Code → Frappe HRMS
Ravi Jangir

The brief: stop hand-entering timesheets; stop pulling monthly project reviews by hand. What followed was a live Microsoft 365 ↔ Claude Code ↔ Frappe HRMS wire-up — no custom code, just existing APIs.

Architecture diagram: Microsoft 365 via MCP, Claude Code as orchestrator, Frappe HRMS via REST API
The architecture, on screen. Microsoft 365 (Outlook, Calendar, Teams chat, Directory) connects via MCP to Claude Code as the orchestrator, which then calls Frappe HRMS via its REST API. No custom code was written. No scripts. Just Claude Code + MCP tools + Frappe REST API.
Ravi’s auto-filled Frappe HRMS timesheet with 6 line items
The output, six minutes later. Six timesheet lines, 8.5 total hours — each one mapped from a real meeting, task or mail that day. Projects auto-matched (Office Admin, Piramal-ESIC, Dhwani Operations, HR Human Resource). Descriptions written from the underlying source.
Ravi’s Project Governance Review page — April review, auto-populated project snapshot
Extension — monthly project governance. Same pattern, different surface. Review Month, POC, Tech Lead, Team Involved, contracted FY revenue, invoiced amount, payment received, collection rate — all auto-populated from the underlying systems. What used to be a manual monthly grind is now a review, not a build.

The Q&A that mattered

Q

“Is our code being used to train the model if we log in with a personal account?” Ravi clarified: with the correct account setup — Dhwani’s Claude Team subscription — your code and prompts are not used for training. The CLI does not process your project data into a training pipeline by default. This is the right answer to give stakeholders who will ask the same thing.

What to steal from this build: the pattern is portable. If you have an internal system with a REST API and you live in Microsoft 365, you can collapse a recurring ops workflow into a two-minute Claude Code conversation. Timesheets is the hook. The real prize is any repeating admin task where the data already exists in two different systems.

02

Kashish Talwani Builder Protocol — ATMA

From never having opened a CLI to shipping an end-to-end Frappe project for ATMA Education Foundation
Kashish Talwani

The honest version of “starting from zero.” Kashish had never used CLI before this engagement — Ankit and Nihaan handed her API keys, a dependency list, and a login page to build. She came back with the rest: verticals, cohorts, NGO partners, MIS admin, web forms, workflows, document review. FRD/BRD markdown straight into live Frappe staging.

Starting point
Zero CLI experience
Shipped
Full ATMA project on staging
Approach
FRD/BRD in markdown → Frappe DocTypes
The Requirements folder showing Atma MIS FRD v1, Verticals & Programme Processes, and supporting references
The input side. A Requirements/ folder with the FRD in markdown and Word, reference images, and reference spreadsheets. Kashish gave AI the context before asking it to build.
ATMA Level-by-Level Programme Flowchart — Organisation to Verticals to Programmes to Cohorts to NGO Partners to Projects
The mental model, first. Before typing a DocType, she mapped the whole hierarchy: Organisation → Verticals (5) → Programmes → Cohorts → NGO Partners → Projects → Process steps.
ATMA MIS Administration on stg.atma.dhwaniris.in — Programmes, Sub-Programmes, Cohorts, Consultants, Web Forms
The live MIS Administration. Master Data with 11 Programmes, 7 Sub-Programmes, 12 Cohorts, 5 Consultants. Web Forms exposed for 22 public links — external NGO registration flows.
ATMA Document Review — Detailed Checklist web form in the live deployment
Web Forms for external users. A live Document Review detailed checklist — the kind of form NGOs fill without needing an internal account. Spec’d in FRD, built via CLI, deployed straight to staging.

What changed between the start and the end

1

Memory as a folder, not a conversation. She used a HEARTBEAT.md + a memory store + a Requirements/ folder with the BRD and FRD in both markdown and Word. Module-wise fields, data types, options, doc references — all there before she asked AI to do anything.

2

Builder Protocol, lived. Not admired. She followed it. Theme → Form → Dashboard, with the FRD as the spec and the CLI as the builder.

3

Direct-to-staging deploys. API keys in the terminal, reference to staging, small iterative changes. No handoff to a dev for minor updates.

4

Credit given: Devang’s shared skill library was the reference that unblocked her. Skills as a library — the organisation tier — actually works when it exists.

The takeaway no one expected: the hard part was not the Frappe code. The hard part was writing the FRD well enough that AI could build from it. Once the spec is crisp, the build is cheap.

03

Arpita Jain Help Center — Welspun

Org-level resources + role-gated FAQs inside a live mGrant deployment — installable in five minutes
Arpita Jain

The problem: clients had an org-level knowledge base need, but mGrant only had project-level file storage — manuals, field photos, creatives, FAQs all trapped inside individual projects. Arpita built the Help Center as a new module to fix that: Resources + FAQs, role-level visibility on folders and files, external link support.

Welspun Help Center — Resources tab with Field Images, Social Media, User Manuals folders
The Help Center home. Two tabs: Resources and FAQs. Three starting folders — Field Images, Social Media, User Manuals. “+ New Folder” with role assignment, “+ Add Resource” for content. The whole surface is new; none of this existed in mGrant before.
Help Center Resources tab with Add Resource button highlighted
Role-level visibility on every folder and every file. When Arpita creates a folder — say CSR — she assigns it to CSR Donor Admin only. NGO users, vendor users don’t see it. Same rule on individual resources inside a folder. This was a direct client ask; it landed in the module.
WelPrakruti field image — a vermicomposting project photo from the Welspun CSR programme
Real content, not placeholder. A WelPrakruti field image from the Welspun CSR programme, uploaded into the Field Images folder. The module handles images, PDFs, Word, plus external links — YouTube, LinkedIn, Drive — so clients can centralise creative and event content alongside documents.

FAQs — the bit ticketing teams will love

1

User manuals drafted with Claude. Arpita generated the first pass of user manuals with Claude, then hand-edited. Those became the seed FAQs, each with screenshots.

2

Inter-page redirection baked in. An FAQ that mentions “My Organization” links straight into that page. “How do I navigate to projects?” goes to the Projects page. FAQs are a navigation layer, not just text.

3

Role-gated FAQs. “How do I delete a user” is visible only to Donor Admins — not to NGO or partner users. Same role model as Resources.

4

Ticketing future. When mGrant ticketing goes live, a raised ticket will first redirect the user to the related FAQs. Self-serve before human.

The deployment story: Welspun took the long way — real build, real iteration. The second client will take five minutes. All of it driven by an MD file that describes the organisation. This is what reusability looks like when the architecture is right from the start.

A Teaching Moment Live

Abhinav, near the 59-minute mark

“When you’re building on an mGrant surface, reference the mGrant design system skill in your project MD before you start. The skill already exists. Make it the default — don’t hand-roll UI from existing project pages when there’s an org-tier skill for the surface.”

Ninety seconds on the call, and the rule of the hour. Skills you don’t know about don’t throw errors — they just cause drift. By the time a reviewer notices, you’ve already shipped.

The rule, written down: if there’s a Dhwani org-tier skill for the surface you’re building on, reference it in your project MD before the first AI turn. This is precisely why an organisation-tier skill library matters — and why Session 6 picks up exactly where this leaves off.

Key Takeaways

1

Four concepts, one stack. Skill = knowledge. MCP = access. Plugin = bundled install. Agent = orchestration. Ship them together, not separately.

2

Dhwani needs its own organisation-tier skill library on GitHub. Coding hygiene, testing standards, design system, domain rules. If you solved it once — contribute the skill. This was the single biggest ask from the room.

3

MCP + existing APIs beat writing new code. Ravi built a production-grade timesheet + governance workflow with no custom code. The win comes from connecting what already exists, not building new services.

4

Spec quality sets build speed. Kashish’s hardest work was writing the FRD/BRD well. Once the spec was crisp, Frappe scaffolding was cheap. The AI executes; you architect.

5

Reusability is an architecture choice, not an afterthought. Arpita’s Help Center was slow for Welspun and instant for the next client because the module was designed to configure from an MD file, not hand-coded per customer.

6

If there’s a skill for your surface, reference it in your MD. Missing skills don’t error — they just drift. The design-system teaching moment made this rule concrete.

7

Your code is not training the model. With the Dhwani Claude Team subscription, your prompts and code stay yours. Say that explicitly to anyone in your project team who asks.

Further Reading

Bonus
Claude Code Session Management & 1M Context — Thariq (16 Apr 2026)

Published the same day as this session. Practical guide to /rewind, /compact, fresh sessions, and subagents now that Claude Code has a 1M token context window. Directly useful for anyone now using CLI long-running sessions — which, after this week, is most of you.

Guide
Equipping Agents for the Real World with Agent Skills — Anthropic Engineering

The official post on Agent Skills. Three-tier progressive loading, why it changes token economics at org scale.

Docs
Model Context Protocol — Official Spec

The open standard that made Ravi’s Microsoft 365 ↔ Claude Code ↔ Frappe HRMS wiring possible. Read the servers section to see what’s already available.

Course
Introduction to Agent Skills — Anthropic Skilljar

Free course from Anthropic on writing your first skill. Start here if the anatomy section above was the part that clicked for you.

Article
When to Use Skills vs. Commands vs. Agents — Daniel Miessler

Practical decision framework when you’re unsure which of the four concepts to reach for. Short. Concrete.