Working With AI · A ground-up session for the team

YOU'VE BEEN USING AI LIKE A CALCULATOR

16 July 2026 Tuhin Patra 
Group Head, Digital
HAVAS, BLR Village

The problem isn't that AI is hard. It's that we're using it like a search bar.

What most of us do
Ask a question → copy the answer → close the tab.
What it's for
A running loop of files, tools, checks, and a finished output.
That's a 5% use of a tool built for 100%.

You'll be able to name every part of the machine - and make one of them work for you

1Decode the vocabulary - tokens, models, context, effort, hallucination, agent
2Know the difference between a skill, a connector, an MCP server, and a plugin
3Understand the four Claude surfaces and when to use which
4Learn the method: goal-based loops, not prompts
5Leave with one asset you own - your own .md file
5 Acts 1 Goal
01 The Machine
02 The Toolkit
03 The Method
04 The Future
05 Your Assets
You can't operate what you can't name. So we start with the names.
Act 01

THE MACHINE

What is actually happening when you type into that box?
01 The Machine
02 The Toolkit
03 The Method
04 The Future
05 Your Assets

An LLM is not a database that looks things up. It's a prediction engine that guesses what comes next.

LLM = Large Language Model
It read an enormous amount of text. It learned the patterns in that text. Given some words, it predicts the most likely next words - one piece at a time.
That's it. That's the whole trick.
The client asked for a 20% reduction in…
CPMRs 120
spendRs 12,00,000
frequency2.5
↑ the model choosing the next token

Your phone's keyboard does this with auto correct.
An LLM just does it with a library instead of a dictionary.

Your keyboard
"The client asked for a 20% reduction in…"theita
Claude

"…the CPM while maintaining a steady daily run rate and frequency”

Same mechanism. Different scale, different training, different depth of pattern.

Not every job needs a specialist. Some jobs need a fast intern who never sleeps.

LLM (Large)
SLM (Small)
Size
Very large
Compact
Strength
Reasoning, nuance, strategy
Speed, cost, narrow tasks
Runs on
Big cloud infrastructure
A laptop or a phone
Media analogy
Your strategy lead
Your trafficking exec
Use it for
"Build me a channel strategy"
"Tag these 8,000 comments"

The model doesn't read words. It reads tokens - and it bills you in them.

"Instamart"Instamart= 2 tokens
"campaign"campaign= 1 token
"₹"₹= 1 token
Rule of thumb: 1 token ≈ 0.75 English words. 100 words ≈ 130 tokens.

You're not billed by the message. You re-pay for the whole conversation - every time you hit enter.

Turn 1
~1k tokens
Turn 10
~12k
Turn 25
~40k
Turn 40
~70k
Every turn re-sends
INyour new message plus the entire history, every file, and the tool list OUTeverything it writes back THINKINGreasoning it burns before answering - invisible, still billed
Long, messy threads get slow, dumb and expensive at the same time. Start clean threads, keep context lean, and put reusable instructions in a .md file instead of re-typing them.

The context window is the desk. When the desk is full, things fall off the edge.

The maximum tokens the model can hold in view at one time. Frontier models today: roughly 200,000 tokens. Some go to a million - a few hundred pages. Big, but not infinite.
What falls off the desk gets forgotten. Not "deprioritised." Forgotten.
the desk · context window
your brief from 40 messages ago

One brand, many engines - and picking the wrong one is how you burn budget

HaikuFast, cheap, lightFormatting, extraction, classification, bulk
SonnetThe workhorse · default80% of daily work. Drafts, analysis, most agent runs
OpusDeep reasonerHard strategy, thorny debugging, multi-step planning
Frontier tierMythos class / Fable 5

State of the art frontier LLMs for deep reasoning, use before 19 July 2026

Numbers after the name = the generation. Sonnet 5 is newer than 4.5. Higher is newer, not always better for you.

Effort is a dial, not a setting you leave alone.

Low effort
Answers fast, from pattern. Cheap.
"Reformat this table."
High effort
Reasons through steps, checks itself, then answers. Slower, more tokens.
"Why did CPM jump 40% in week 3?"
You'd give a junior 5 minutes for a formatting fix and 3 days for a channel strategy. Same instinct. Same dial.

Every AI word is an ordinary word in a costume. Here's the costume remover.

The word
What it actually means
In plain office language
Prompt
What you send it
The brief
Token
A chunk of text
The unit on the meter
Context window
What it can hold in view
The desk
Model
The trained brain
The person you assigned it to
Effort / Thinking
How long it reasons first
Deadline pressure
Temperature
Predictable vs. varied wording
Safe copy vs. wild copy
System prompt
Standing instructions
The job description
Inference
The act of generating an answer
Doing the work
Fine-tuning
Retraining a model on your data
Sending it to your academy
RAG
Fetching your docs as context
Handing it the file first
Agent
A model that loops: acts, checks, retries
An employee, not an advisor
Hallucination
Confident, fluent, wrong
The intern who won't say "I don't know"

It never lies to you. Lying requires knowing the truth.

Hallucination = the model produces something fluent, confident, formatted correctly - and false. It isn't malfunctioning. It's doing exactly what it was built to do: predict plausible text.
Plausible ≠ true.

It's a jazz musician, not a librarian. You asked for a song and it will always play one.

The librarian
Asked for a book that doesn't exist: "We don't have that."
The jazz musician
Asked to play a song they don't know: plays something in the right key that sounds exactly like it.
An LLM is the jazz musician. Fluency is the instrument. Truth is not.
·Invents a plausible CTR benchmark, sourced to a report that doesn't exist
·Cites a case study - right brand, right year, wrong numbers
·Confidently describes a platform feature deprecated last year
·Gets 9 rows of a table right and quietly fabricates the 10th
Where SLMs are worse: smaller model, thinner patterns, more confident nonsense on anything niche.

You don't fix hallucination with better prompts. You fix it by removing the need to guess.

If it's guessing about…
Give it…
Your client's numbers
The actual export. Attached.
Market data
A search tool, and demand the link
Your process
A skill file that spells it out
Anything it can't verify
Permission to say "I don't know"
The rule: every number in a client-facing output traces to a file or a link. No exceptions.
What is an agent?

A chatbot answers. An agent gets it done.

007
agent, not analyst
A software system that uses a model to pursue a goal and complete multi-step tasks on your behalf - it reasons, plans, remembers, uses tools, and acts with autonomy.
Goal
"Get the intelligence." The mission, not the method.
Tools
The gadgets - things it can actually do in the world.
Memory
The briefing and past missions it carries in.
Connectors
MI6, safe houses, local assets - keys to other systems.
Its job is to deliver the result - not to lecture you. Act 2 is how we hand our agents their gadgets.
Act 02

THE TOOLKIT

A brain alone is useless in an office. Here's what we bolt on.
01 The Machine
02 The Toolkit
03 The Method
04 The Future
05 Your Assets

The model is just the brain. The stack around it is the whole machine.

The modelthe brain
+ Contextwhat's on its desk right now
+ Skills & pluginshow we do things here
+ Connectors & MCPkeys to the other rooms
+ Memorywhat it remembers about you
+ Computer usehands, for when there's no key
= Something useful

Everything in this act is = onboarding a new employee

The layer
The onboarding equivalent
Model
The person you hired
Context
The brief you handed them this morning
Skill
The SOP binder - "this is how we build a PCA here"
Connector
Their Gmail login, their Drive access
MCP server
The IT plumbing that makes those logins possible
Plugin
The onboarding kit - binder + logins + tools, one box
Memory
What they remember about you from last quarter
Computer use
Them clicking the mouse when there's no API

A skill is your process, written down once, executed forever

A skill = a folder with a SKILL.md file inside, plus any reference files it needs. It tells the model how you do a specific thing - your standards, structure, format, rules.
The model reads it automatically when the task matches. You write it once. It never forgets it. It never gets it 80% right.
📁 pca-skill/
SKILL.md
Same input →
✓ identical output
✓ identical output
✓ identical output

I didn't teach it media planning. I taught it our media planning.

media-plan
Paste a raw brief, get a structured digital plan. Never recommends TV or OOH - our brief is digital-only. It knows because I told it once.
deck-architect
Knows our narrative spine: Reality → Intelligence → Strategy → Activation → Proof. Knows headlines are conclusions. Knows the forbidden phrases.
cpm-diagnostics
CPM spikes - it isolates the cause: fatigue, bid landscape, creative decay, LP drop, budget shift, new competitor.
Total build time each: under an hour. Total times reused: forever, it's self learning, self healing.

A connector is a key to a room Claude couldn't otherwise enter

A connector = a ready-made link between Claude and an app you already use. Turn one on → Claude can read from and act in that app.
Gmail Google Drive Calendar Notion Slack Figma Canva Shopify Supabase Meta Ads Similarweb Zoho Books
Turned onNot yet
Without a connector, Claude is a very smart person with no logins.

MCP is the USB-C of AI tools - and it's why your connectors exist at all

MCP = Model Context Protocol. An open standard for how AI models talk to tools.
Before MCP
Every AI × every tool = a custom, one-off integration.
tangled spaghetti · chaos
After MCP
Build one MCP server - every AI model that speaks MCP can use it.
one clean port
A connector is the button you press. An MCP server is the plumbing behind the wall. Anyone can build one.

A plugin is the onboarding kit - skills, tools, and commands in one box

A plugin = a bundle. Skills + connectors + commands, packaged so someone else can install it in one click. You build a way of working. You zip it. Your whole team runs it.
📁
Skill
🔑
Connector
/
Command
This is how a personal workflow becomes a team standard - and eventually, an agency asset.

Skills teach. Connectors access. MCP connects. Plugins distribute.

What it is
Answers
Analogy
Who makes it
Skill
Instructions in a file
How do we do this?
SOP binder
You
Connector
A live link to an app
Can you reach my data?
Office key
The platform
MCP server
The open protocol underneath
How do tools plug in at all?
USB-C port
Any developer
Plugin
A bundle of the above
Can the team have this too?
Onboarding kit
You, then shared

Memory is the difference between a consultant and a colleague

Without it
Every session starts at zero. You are a stranger, every single day.
With it
It knows your clients, your formats, your tone, your standing rules.
Context = this desk, today.Memory = what it knows about you, always.
You control it. Add, edit, delete. Say "forget that" and it's gone.

When there's no key to the room, it learns to use the door handle

Computer use = Claude sees your screen, moves your cursor, clicks, types. No API. No connector. No integration. Just the same interface you use.
Why it matters
Most software in our lives has no AI connector and never will. Computer use is the universal fallback.
Why it needs care
It can see whatever's on screen. Screenshots included.

The browser stops being something you drive and becomes something you delegate

1Chrome Web Store → search Claude → verify the publisher is Anthropic
2Add to Chrome → Add extension
3Click the puzzle-piece icon → pin Claude to your toolbar
4Open the side panel → sign in with your paid Claude account
5Grant site permissions - start with "ask before acting"
Paid plans only. Chrome and Edge. Beta.

It doesn't summarise the page. It works the page.

Navigate, click, fill forms across tabs Record a workflow once - repeats it forever Schedule browser tasks: daily, weekly, monthly Drag tabs into a group → reads all of them together Reads console errors & network requests
One security rule, non-negotiable
Websites can hide instructions inside a page to hijack your agent - it's called prompt injection. Start with sites you trust. Never let it touch banking, passwords, or anything irreversible without you watching.

Same brain. Four bodies. Pick the body that fits the job.

Surface
What it is
You use it for
Claude (chat)
The box you already know
Thinking, drafting, asking
Claude in Chrome
Agent in your browser
Anything that lives on a website
Claude Code
Agent in your files & terminal
Building things. Not only code.
Claude Cowork
Agent on your desktop, for non-devs
Real deliverables from real files
Claude Dispatch
Your phone, controlling Cowork
Work that continues while you don't

Claude Code. It is not only for code.

An agent that lives where your files are. It reads them, edits them, runs things, checks its own work, and iterates - for hours if you let it.
·Build an internal tool nobody will fund but everyone needs
·Vibecode a microsite for a pitch instead of describing it
·Batch-process 40 platform exports into one clean sheet
·Run a repeatable analysis the same way, every week
The barrier is real: it looks like a 1980s terminal. That's a design problem, not a capability problem.

Cowork is Claude Code with the terminal removed - which is to say, it's for us

Claude Cowork = an agent on your desktop, built for knowledge workers instead of devs.
Your local files Your connectors Your skills Your plugins Your browser Computer use
Messy folder of CSVs, a brief, a rate card → A formatted PCA deck
The terminal was the tax. Cowork removed the tax.

Your phone becomes the remote control for the agent sitting at your desk

Dispatch = message Claude from your phone; Cowork does the work on your desktop.
·One continuous thread. Pick up where you left off, phone or laptop.
·Uses everything your desktop has - files, connectors, plugins, apps.
·Your desktop must be awake with Claude Desktop open. Not a cloud service.
·Beta. Pro and Max plans. Needs both the desktop and mobile apps.
Setup: Desktop app → Dispatch in the left panel → Get started → toggle file access and keep-awake → start messaging from your phone.

If you're asking, use chat. If you're assigning, use Cowork. If you're building, use Code.

Do you want an answer, or a deliverable?
↓ ANSWER
CHAT
↓ DELIVERABLE
Files on your desktop, or on a website?
DESKTOP
COWORK
away from desk?
DISPATCH
WEBSITE
CHROME
APP / WEB APP
CODE
a tool, not just a file
Act 03

THE METHOD

Knowing the parts is not the same as knowing the move.
01 The Machine
02 The Toolkit
03 The Method
04 The Future
05 Your Assets

PROMPTING
IS DEAD

Prompting isn't dead. One-shot prompting is. The skill was never the sentence - it was the setup.

The old game
Craft the perfect paragraph → get the perfect answer → copy it out. It worked because the model was weak and had no tools.
The new game
The model is strong and has tools. What limits it now is not your wording. It's what you gave it and what you let it do.
Prompt engineering → context engineering.
Context engineering sets up what it has. The loop - next slide - is what it does with it.

Stop writing requests. Start setting goals and letting it run.

GOALwhat "done" looks like, and how we'll know
↓
PLANit proposes the approach. You approve.
↓
ACTit reads files, calls tools, writes
↓
OBSERVEit looks at what came out
↓
CHECKagainst the goal. Not done? Loop.
↓
DELIVERthe artefact. Not a chat message.
↻
Miss the goal? It goes back to ACT and tries again - on its own. That loop, running without you, is the whole shift.
You define the goal and the guardrails. It runs the loop. You inspect the gates.

The difference between an instruction and a brief

✕ PROMPT
"Write me a PCA summary for the campaign."
✓ GOAL
Build the PCA for [Client], following our pca skill and clients/[client].md.
Sources: the four platform exports in /downloads/campaign-oct/ and last campaigns PCA for the trend line. Every number traces to one of those - nothing inferred.
Done means: delivered vs. plan, the three findings that actually move the conversation (each backed by a chart), and two recommendations with a projected delta and the assumption behind it.
Guardrails: no benchmark you can't link. Where the data's thin, say so rather than smoothing it over. Leave the client's campaign naming exactly as-is.
Check yourself against the goal before you hand it to me.

Sources. Done. Guardrails. Check. If your goal is missing one, it will show up in the output.

Sources
Where the truth lives. Files, links, tools. Nothing invented.
Done
The definition of finished. Testable, not vibes.
Guardrails
What it must not do. The forbidden list.
Check
How it verifies itself before you see it.
Token maxing

Token maxing isn't using fewer tokens. It's refusing to spend them twice on the same thing.

Spend tokens on
The actual reasoning
The retry loop
Reading your real data
Self-checking the output
Never spend tokens on
Re-introducing yourself
Re-pasting the same brief
Correcting a mistake you fixed last week
Re-typing your formatting rules
Bad: cramming everything in and hoping.Good: exactly what it needs, nothing it doesn't.

The most powerful file format in AI is a plain text file from 2004

Markdown (.md) = plain text with a few marks that mean structure.
That's the entire language. You just learned it. It took nine seconds.
# A heading
## A sub-heading
- A bullet
**Bold for emphasis**
| A | Table |

A Word doc spends your tokens on fonts. A PDF spends them on layout. A markdown file spends them on meaning.

Format
What the model wades through
.docx
XML, style definitions, revision metadata, font tables
.pdf
Layout coordinates, text fragments, broken reading order
.pptx
Shape positions, theme data, slide masters
.md
Your words. The structure. Nothing else.
Same information. A fraction of the tokens. And the headings tell the model what matters - hierarchy for free.

Your prompts are disposable. Your .md files are equity.

A prompt dies when you close the tab. An .md file:
Compoundsimproves every time you use it
PortsClaude today, any model tomorrow. Nobody owns it.
Distributeshand it over; they get your standard, not your effort
Survivesyou leave the team; the standard doesn't
This is the closest thing to intellectual property an individual knowledge worker can own.

Five files. That's the whole starter kit.

File
What's in it
CLAUDE.md
Who I am, what I do, my clients, my tone, my defaults, my rules
clients/[brand].md
Category, audiences, history, sensitivities, naming, what they hate
formats/pca.md
benchmarks.md
Our numbers, with sources and dates. Never guessed.
never-do.md
The forbidden list. Every mistake, written down once, never repeated
Start with one. CLAUDE.md. Twenty lines. Tonight.

The agent does the work. You carry the risk. That asymmetry is why you stay in the loop.

Human in the loop = a human approves at the moments that matter.
·It is confidently wrong sometimes - and it sounds identical to when it's right
·It has no stake. If the number is wrong, your name is on the deck
·It optimises what you measured, not what you meant
·Actions have consequences it doesn't feel. Spend. Send. Publish. Delete.
The rule: read-only by default. Approval before anything that leaves the building.

Not every step needs you. Four of them absolutely do.

Gate
The question
Skip it?
Goal
Is this the right thing to do?
Never
Plan
Is this the right way to do it?
Rarely
Facts
Does every number trace to a source?
Never
Send
Does this leave the building?
Never
Format
Are the fonts right?
Yes. Let it go.
Spend your judgment where judgment is scarce.
Act 04

THE FUTURE

Where this is going, and why it changes what "work" means.
01 The Machine
02 The Toolkit
03 The Method
04 The Future
05 Your Assets

You're not getting an assistant. You're getting a staff.

A personal AI agent = an agent that knows you - your files, accounts, standards, calendar - and acts on your behalf, on your instruction, at your risk.
Not a chatbot. A chatbot answers.An agent goes and does, then comes back.
Dispatch is the earliest recognisable version of this. Your phone. Your desk. Your work, continuing while you're in a meeting.
Think about it like this

The UI exists because humans were the only things clicking. That assumption is expiring.

Remember the telephone switchboard?
Two people wanted to talk. But the wires couldn't connect themselves, so we hired an operator to plug them in by hand. The operator was never the point - she was the workaround for wires that couldn't route themselves.
The interface is that operator
Every form, dashboard and portal is a human plugging wires by hand - because software couldn't talk to software. Now it can. So who needs the switchboard?
Take the human out, and the interface is just friction between two machines.
Today, all a workaround: Airline sites Hotel forms Ad platform UIs Vendor portals RFP templates

You describe the campaign once. Your agent clicks through Meta so you don't have to.

The task nobody wants: set up an auction campaign on Meta
"Set up the always-on prospecting campaign - 8 ad sets, our standard audiences, ₹5L/day, the creatives in this folder."
Today - 90 minutes of clicking
Build each ad set by hand. Re-pick the same audiences. Re-type budgets. Upload creatives one by one. Fix the two you named wrong. Every single time.
Next - one sentence, then you check it
Your agent talks to Meta's MCP server, builds all 8 ad sets to your naming convention, and stops - paused, ready for your review before anything goes live.
You still hit publish. But the 90 minutes of identical clicks is gone.

This is the direction. It is not the date.

Works today
Agents that read your files, use your tools, and run real tasks - Cowork, Dispatch, computer use. You can do this on Monday.
Not yet
Handing an agent your card and walking away. Trust, payments and security aren't solved - keep a human at the gate.
Build the muscle now.

If the portal disappears, so does the job of working the portal

Work that gets absorbed
Work that gets more valuable
Pulling audience numbers out of Google Reach Planner
Deciding which audience is worth buying
Exporting the raw dump from Google Keyword Planner & Meta Audience Manager
Reading the intent hiding in that data
Plan-vs-delivery reports via the reporting-skill.md
Explaining why delivery drifted
Building the plan with media-plan.md + rate cards RAG'd from Drive
The judgment call on where the budget actually goes
Assembling the PCA from this morning's Google Ads, DV360, Meta & manual Sheets
The story the client actually needs to hear
Nobody is being replaced by AI. People who operate it will replace people who don't.

If one person can do the work of five, the question stops being "how many hours?" and becomes "how much output?"

The provocation
Compensation tracks leverage - output per person - not hours. "I ran 40 agent hours and shipped six decks" means something on a review form.
The reality check
Nobody is paid in tokens today. No company has this system. It's a thought experiment, not a forecast.
Already true
The gap between the top operator and the average one just widened by 5×. That gap always gets priced.

The asset isn't the tokens. It's the file that makes the tokens worth spending.

Tokens are fuel. Anyone can buy fuel.
The skill file, the client file, the format file - that's the engine.
One has a library
Same budget, same model.
One has a chat window
Same budget, same model.
It isn't close.
Act 05

YOUR ASSETS

Four things to build. All of them yours. None of them need permission.
01 The Machine
02 The Toolkit
03 The Method
04 The Future
05 Your Assets
Asset One

The file that means you never explain yourself twice

Start with CLAUDE.md. Twenty lines. Tonight.
Then add one line every time you correct it. That's the whole practice.
# Who I am
Group Head, Digital. Havas Media India, Bengaluru.
Media strategy, planning, buying, AdTech.
# My clients
[Brand] - [category, what matters, what to never do]
# How I work
Boardroom-ready or don't bother. Headlines are conclusions.
Every number traces to a file or a link. No exceptions.
Digital only. Never TV, OOH, print, radio, cinema.
# Never
Fabricate a benchmark. Invent a case study. Guess numbers.
Asset Two

Turn the thing you do every week into a thing you did once

How to find your first skill - one question:
What did I do this month that I'll do again next month, the same way?
PCA structure Competitive scan Brief-to-plan Rate card cleanup Weekly pacing check
Write down how you do it. Save as SKILL.md. You're done.
Platform-agnostic. Works with Claude today and whatever replaces it in three years. Nobody can take it off you.
Asset Three

Build the tool nobody will ever fund but everyone needs

The tool that only solves your problem was never worth a dev sprint. Now it costs an afternoon.
A pacing checker that eats four platform exports and flags under-delivery
A rate-card normaliser that makes six publisher formats into one grid
A creative-fatigue watcher that pings you at frequency 4
A brief-to-plan starter that outputs your first cut in the house format
Give it a simple login and a link, and the same tool becomes something the client logs into - a live pacing dashboard, a self-serve rate estimator. You stop sending weekly screenshots and start owning the tool they open every morning.
Asset Four

Stop describing the idea on a slide. Hand them the thing.

Vibecoding = you describe it in plain English, the agent builds it, you react, it iterates.
Slide 14 says "an interactive dashboard for your always-on brand health"
→ forgettable
You open the working dashboard on their data
→ they lean in
Microsites. Prototypes. Interactive PCAs. Live tools. Built in an afternoon.
The highest-leverage thing on this list. A working prototype changes the temperature of a room in a way no slide ever has.

Before any of this: we are custodians of other people's businesses

-Client data goes into approved tools only. Check before you paste. If unsure, ask.
-Read-only by default. Approve before anything sends, spends, publishes, or deletes.
-Every number in a client-facing output traces to a source. Fabrication is not a mistake - it's a breach.
-Computer use sees your screen. Close what shouldn't be seen.
-Prompt injection is real. Untrusted websites can hijack your agent.
-You are accountable for the output. Always. The tool is never the excuse.
The Demos · 25 minutes

Enough theory. Let's watch it actually work.

Three live demos - each one a task you do by hand today:
1
Set up a Meta campaign
End to end, in Cowork
2
Competitive ad intel
AdSignal - a tool we built
3
Pull reach numbers
Google Reach Planner, via computer use
Something might break live. That's fine - that teaches more than a perfect run.
DEMO 1 · VIA WHATSAPP · BUILT USING CUSTOM MADE AGENTS

Set up a full Meta & Google campaign end to end - from one sentence or maybe a voice note

1One instruction: "Build the always-on prospecting campaign - 8 ad sets, our standard audiences, ₹5L/day, creatives in this folder."
2The agent hatches, opens Ads Manager and works it like you would - and setup the campaign in seconds
3Builds all 8 ad sets to our naming convention, sets budgets, picks the saved audiences
4Uploads the creatives, matches each to the right ad set
5Then it stops. Everything staged, nothing live.
6I review, fix the one thing I'd do differently, and hit publish.
Watch for
It driving the real Ads Manager UI · the naming convention held perfectly · the hard stop before anything spends
90 minutes of identical clicking, done in the background. I still own the publish button.
Demo 2 · AdSignal · a Chrome extension we built

Competitive ad research that ate a day - now runs in under a minute

The pain it kills
Every competitive review means hand-trawling the Meta Ad Library and Google Ads Transparency Center - scrolling hundreds of ads, screenshotting rivals one at a time, pasting them into a deck. A full day of grunt work per brand, and it's stale the moment you finish.
AdSignal turns that day into a minute - it does the scrolling, capturing and formatting, so the team keeps the part that actually matters: the read on what competitors are doing.
The ads it captures for you
AdSignal - turning ad libraries into actionable intelligence
Once you understand the machine, you stop just using tools - you build the one you always wished existed.
Demo 3 · computer use

Pull reach & audience-size numbers out of Google Reach Planner - no export, no API

1One instruction: "Get me reach and audience size for these three plans in Reach Planner."
2Reach Planner has no clean export - so it uses computer use, reading the screen like you do
3Sets each market, budget and format; reads the reach curve off the UI
4Captures the numbers into a clean sheet - one row per scenario
5Feeds straight into the Demo 2 plan - no re-keying
Watch for
It reading numbers off a tool with no API · the same clicks you'd make · clean data out the other side
If you can do it on your screen, it can do it on your screen. That's the whole idea of computer use.
The whole session

Forget everything else. Keep these seven.

1
It predicts, it doesn't know.
Every strength and failure flows from that one fact.
2
Fluent ≠ true.
A jazz musician, not a librarian. Ground it or verify it.
3
The model is the smallest part.
Context, skills, connectors, memory - that's the leverage.
4
Stop prompting. Start briefing.
Sources. Done. Guardrails. Check. Then let the loop run.
5
Tokens are fuel. The .md file is the engine.
Write it once. Spend it never again.
6
Your .md files are your IP.
Platform-agnostic. Portable. Compounding. Yours.
7
You stay in the loop - at the gates that matter.
Goal. Plan. Facts. Send. Your name is on the deck, not its.

Do one thing tonight. Not seven. One.

Open a text file. Name it CLAUDE.md. Write twenty lines about how you work.
Everything in this deck is downstream of that file existing.

Four weeks from user to operator

Week 1CLAUDE.md exists. Chrome extension installed. Every correction goes in the file.
Week 2One skill written - the thing you do every week. Two connectors on.
Week 3One Cowork task, scheduled, running without you.
Week 4One thing built you'd actually show someone - a tool, a microsite, a prototype.
Week 5: you're not the same operator.

The tool isn't the advantage. Everyone has the tool.

The advantage is the library you build on top of it.
Start today. Twenty lines.
Tuhin Patra 
Group Head, Digital
HAVAS, BLR Village
Appendix · Q&A prep

The questions that always come up

Q
Will this take my job?
The boring half, yes - and it should. Judgment, relationships and taste get more valuable. But only for the people who move now.
Q
Is it safe to put client data in?
Approved tools only. Check policy first. When in doubt, ask before you paste - not after.
Q
How do I know it's not making things up?
Not by reading harder. By grounding it in a file and demanding a source for every number. Verification is architecture, not vigilance.
Q
Isn't this expensive?
Compare it to the loaded hourly cost of the two hours it just took off your morning. Then ask again.
Q
Do I need to learn to code?
No. You need to learn to write clearly. That was always the actual skill.
Q
Why Claude and not [X]?
It's what I use daily and the surfaces are most complete. Every concept here is universal - your .md files work everywhere.
1 / 72