CLASS 11June 30 · Data Cleanup & Prep for Agentic Workflows

Class 11: Data Cleanup and Prep for Agentic Workflows. "Scale scales whatever systems already exist. AI scales chaos."

Taught byEkta ModyHead of Creative at a Google-brands agency
Ex-Accenture innovation practice, now teaching Creative AI

The class in brief

Ekta's whole night argued one shift: stop thinking in prompts, start thinking in systems. Her tool for that shift is the System Nucleus, a six-part reusable context block. After this page you will be able to do three things.

Build a System Nucleus for one category of your own repeated work.
Audit a real workflow for where handoffs lose context.
Place a review checkpoint where the risk actually is.

Prompts are temporary, systems scale

Why this matters · 6:10 PM

AI doesn't fix broken workflows. It exposes them.

A clear system scales cleanly. A messy one gets amplified just as fast, per . Ekta opened with four learning objectives, shown on , then made that comparison the core argument of the night.

The shift underneath everything that followed: prompts are temporary, systems scale. That means reusable structures and defined roles, instead of rewriting instructions in every new chat.

workshop blocks: System Nucleus, Workflow Architecture, Operational Systems, Review & Scale. Two exercises run across them.

formal graded homework. Ekta left a reflection instead: if you built one system next week, what would it be.

The framework · 6:19 PM

The six parts of the System Nucleus

Goalwhat you're trying to accomplish

The north star for this category of work, not this one project.

Your Rolewhat decisions you own

Names the calls that stay human before AI ever enters the loop.

AI's Rolewhat it helps with

Where the machine actually contributes, stated as plainly as your own role.

Voice / Stylehow outputs should sound

Where brand voice or personal voice lives, so it stops getting re-explained.

Decision Logicthe principles guiding calls

The reasoning a teammate, or a model, should apply when a judgment call comes up.

Sourcesreferences AI should always consider

The material that should never need re-uploading.

The slide's five-part nucleus, fully labeled — The slide lists five parts. Ekta's own worked example splits Roles into Your Role and AI's Role, which is why the six-row template on this page follows that same split.

Definition to keep, from Ekta directly: "an operating system for yourself, based on which all similar types of projects can be run." Save it anywhere reusable: a Google Doc, a Claude Project, or a Gemini Gem, per .

The exercise · 6:22 PM

Build the nucleus and then find where it breaks

Exercise 1

Pick a category of repeated work. Fill the six rows, one sentence each.

Choose a whole category, such as client presentations, rather than one project such as the Nike deck, per . Save it in a Google Doc, Notion, or a Claude, ChatGPT, or Gemini project.

After the break, the room mapped where those same workflows actually fail.

The chain that loses context

Brief, research, AI exploration, design development, presentation: Ekta built the chain step by step on screen, then named the exact failures between each link. Context lost between steps, feedback trapped in chats, files hard to find, decisions never captured, outputs that can't be reused. Her line stuck: "AI didn't fail. The workflow couldn't support it."

Data hygiene

Data hygiene means being deliberate about what you feed AI, and what you don't want it to know. "Context is the most important word." Some people design their systems on purpose; most build them by accident, and accidental systems are the ones that break under AI.

One source of truth and a naming formula

The craft · 7:27 to 7:41 PM

Store context once. Reference it everywhere. Name it so a machine can find it.

Operational Systems opened with the trap most people build without noticing: . The fix is one shared project context, whether that is a ChatGPT Project, Claude's Project Knowledge, or NotebookLM, shown on .

Class 4's GOLD framework (Goal, Output, Limitations, Data) got reused here as the saved project brief, so the prompt collapses to one line. Then came naming: a show of hands for who has a file called Final_v2 or Final_FINAL, followed by a four-question fix.

questions a good file name answers: When? Project? Type? Status? Formula: [Date]_[Project]_[Type]_[Status].

Review decisions, not deliverables

The judgment · 7:49 PM

Review is important for decisions, not deliverables.

Ekta polled the room before answering : given a project brief, research summary, ten concept directions, a selected direction, and the presentation, where should one review checkpoint go?

The room's own answer was right, per : right after AI exploration, the moment you decide which ideas move forward. Review direction, prioritization, quality, and alignment there, not every draft, prompt, file, or revision downstream.

checkpoint, placed where the risk actually is, beats reviewing every single output along the way.

Methods and prompts

Five methods to take with you

METHOD 01 · 6:22 PM part of: build tools not chats

Build a System Nucleus

Fill six one-sentence rows for one category of your repeated work, then save it somewhere reusable. This is the block you stop re-explaining.

Where it came fromTaught by Ekta Mody in class 11 as the first exercise of the night, filling six one-sentence rows, Goal, Your Role, AI's Role, Voice or Style, Decision Logic, and Sources, for one category of repeated work.Use it whenUse this when you keep re-explaining the same project context in every new AI chat and want to save it once instead.

Same idea elsewhereStand up a reusable Project or Gem (class 4) · Weight your references on purpose (class 5) · Let one model write the prompt for another (class 6)

Working prompt

Help me build a System Nucleus for [category of work, e.g. client presentations]. Ask me one question at a time for each: Goal, Your Role, AI's Role, Voice/Style, Decision Logic, Sources. Keep every answer to one sentence, then compile it into a block I can paste at the start of any related project.

You will know it worked whenit asks about each of the six fields one at a time, then hands back a single compiled block you can paste at the start of any related project.

METHOD 02 · 7:04 PM part of: reject the first draft

Find where the workflow breaks

Name the weak handoff in your own process before asking AI to check your read. The point is naming the break, not getting a flattering audit.

Where it came fromTaught by Ekta Mody in class 11's second exercise, where students named their own guess at the weak point in a workflow before asking AI to map the actual handoffs and check that guess.Use it whenUse this when a workflow that uses AI keeps producing weak results and you want to find exactly where context or feedback gets lost.

Same idea elsewhereSet up a debate partner (class 4) · Build cheap and then swap the model up (class 8) · Critique mode (class 10)

Working prompt

Here is a workflow I run: [list the steps]. My own guess at where it breaks: [your answer, and why]. Now you check me: map the actual handoffs between these steps and tell me where context, feedback, or a decision most likely gets lost, and whether my guess was right.

You will know it worked whenit maps the actual handoffs between your steps and tells you plainly whether your own guess about where it breaks was right.

METHOD 03 · 7:29 PM part of: context beats prompts

Store context once

Stop rebuilding the same brief in every new chat. Save it once as project knowledge and reference it, instead of re-pasting it.

Where it came fromTaught by Ekta Mody in class 11 as the fix for rebuilding the same brief every session, saving a full project brief using the GOLD structure once and referencing it going forward.Use it whenUse this when starting a project you know you will return to repeatedly, so you never have to re-paste the same brief again.

Same idea elsewhereControl the input and build gradually (class 2) · Show it your voice and not just the output (class 3) · Evolve the reference instead of restarting (class 5)

Working prompt

Here is my full project brief, using GOLD: Goal [ ], Output [ ], Limitations [ ], Data [ ]. Save this as the reference for everything else in this project, and every time I ask for something, check it against this brief before you answer.

You will know it worked whenlater requests in the same project get checked against the brief you saved, without you having to paste it in again.

METHOD 04 · 7:41 PM part of: build tools not chats

Let AI name your files

Feed the four-question naming formula in once, then let the tool name its own outputs going forward. Don't fix the backlog, just set the convention from here on.

Where it came fromTaught by Ekta Mody in class 11 as part of her naming formula, feeding a tool a four-part naming rule once and letting it apply that formula automatically to every file it generates from then on.Use it whenUse this once you have settled on a naming formula and want your files named consistently without doing it by hand each time.

Same idea elsewhereStand up a reusable Project or Gem (class 4) · Weight your references on purpose (class 5) · Let one model write the prompt for another (class 6)

Working prompt

From now on, name every file you generate for me using this formula: [Date]_[Project]_[Type]_[Status]. Ask me for the project and type once at the start of a session, then apply the formula automatically to every output without asking again.

You will know it worked whenevery file it generates from here on carries the date, project, type, and status formula, applied without you asking each time.

METHOD 05 · 7:49 PM part of: keep the judgment human

Place the review checkpoint

Decide where your own review checkpoint belongs before asking AI to check the placement. The right answer is usually right after the machine has generated the widest range of options, not after every step.

Where it came fromTaught by Ekta Mody in class 11, where the class was polled on where a single review checkpoint should sit in a multi-step workflow and the room correctly placed it right after AI exploration.Use it whenUse this when designing a workflow and deciding where the one point of human review should go to catch the highest-risk decision.

Same idea elsewhereJudge AI output with your own trained eye (class 1) · The sea-of-sameness check (class 2) · Pick the relationship on purpose (class 2)

Working prompt

Here is my workflow: [list the steps]. My call on where the one review checkpoint should go, and why: [your answer]. Now you check me: does that step carry the highest risk of a wrong direction, and would moving it earlier or later change what gets caught?

You will know it worked whenit tells you plainly whether your chosen checkpoint sits at the highest-risk step, and whether moving it earlier would catch more.

Audit your workflow

The close · 7:56 PM onward

Get your house in order before you build agents

Exercise 2 turned the room's own nucleus into a test subject . Paste your nucleus and a mapped five-to-ten step workflow into a fresh chat with an "AI Workflow Consultant" audit prompt. Then get bottlenecks and weak points back as a table .

The takeaway was never the audit output itself. It was writing the fixes back into the nucleus so it improves for next time.

Ekta's closing slide named the whole night in five lines, per .

Build a nucleus.
Store context once.
Name consistently.
Review decisions.
Improve systems, not prompts.

Next workshop was agents. Tonight was laying the foundation before agents use it.

Try this prompt

Quiz me on the six parts of the System Nucleus and Ekta's naming formula. Then give me a messy example workflow and make me place the one review checkpoint correctly.

You will know it worked whenit quizzes you on the six parts of the System Nucleus and the naming formula first, then makes you place the one review checkpoint correctly in a messy example workflow.

The shelf

Tools and references

Tools that night

Claude, ChatGPT, and Gemini, for building and auditing the nucleus (different tools gave different audit results)
Google Doc, Claude Project, or Gemini Gem, as the save location for a nucleus
The Miro board: System Nucleus templates and the workflow-audit prompt

Named in the room

Figma's Config releases, reported in the AI-news opener: Weave, animation-to-code, an AI plugin builder
Ekta's mentorship-platform list, which she offered to share
"It's never too late to start": don't fix the backlog, set the convention going forward

Every frame from this class

26 captures, in order. Click any one to see it full size.

Prompts are temporary, systems scale

A tan and black table listing Goal, Your Role, AI's Role, Voice/Style, Decision Logic, and Sources for a client presentation. — Ekta's worked example fills in a full System Nucleus for a client presentation. It shows the goal, who owns which decisions, what AI should help with, the voice, the decision logic, and the sources to pull from.

Title slide reading Data and File CleanUp for AI Workflows on an orange to blue gradient background, with a thumbnail slide rail on the left. — The class title slide for Data and File CleanUp for AI Workflows, shown with the deck's slide rail visible as Ekta begins her presentation.

Ekta Mody, the class's guest speaker, is Head of Creative for Google and other global tech brands, with 15+ years in human centered design and a background leading Accenture's global design practice.

Tony Jones, cofounder of Creative AI and instructor of the AI Design course at Pratt, has 20 years in advertising at agencies including Anomaly and McCann.