How We Built ‘Claudie,’ Our AI Project Manager (Full Walkthrough)

Every’s Head of Consulting Natalia has spoken to over a hundred companies in the past year about AI adoption. In this conversation with Dan Shipper, she distills what separates companies that get real leverage from AI from those where it fizzles into a few power users and a lot of floundering. The core insight: AI adoption is a coordinated effort requiring both top-down leadership and empowered internal champions — and generic prompting fails in specialized industries. The second half is a live technical walkthrough of “Claudia,” an autonomous AI project manager built in Claude Code that reduced Natalia’s weekly project management overhead from 10–15 hours to roughly one hour.

The conversation also surfaces an underappreciated pattern: the people who build the most effective AI systems are not pure engineers. They’re domain experts paired with technical partners, iterating through multiple failed approaches until they find the architecture that captures both structural rigor and deep knowledge of what “good work” actually looks like.

What Works: Lessons from 100+ Companies

Every’s consulting practice works with hedge funds, PE firms, Fortune 500 companies, and tech organizations on AI strategy and implementation. After speaking with over a hundred companies in the past year, Natalia identifies two patterns that consistently separate the companies getting real leverage from AI.

1. It Has to Come from the Top Down

AI adoption is unlike historic software rollouts. You can’t just let the CTO buy a tool and hope people use it. The companies that go furthest are the ones where the CEO is personally engaged — not just mandating AI use, but actively experimenting with it.

Dan reinforces this: “You will probably go as far in terms of AI adoption as your CEO has gone. It’s not something the CEO can delegate.” He points to Tobi Lütke at Shopify as a public example — a CEO who hacks on AI projects on weekends, which fundamentally shapes the company’s culture around the technology.

2. Empower Internal Champions

Inside any organization, there are natural early adopters. The executive’s job is to identify these people, elevate their status, and spread what they know. These “AI Champions” need creative power to rethink their roles, experiment, fail, and then double down on what works.

This requires something counterintuitive: giving people space to play. Every holds a “Think Week” every six months where the team drops all day-to-day work and just experiments with new technology. Natalia and her applied AI engineer Natash Agarwal started meeting at 6:00 a.m. daily, three hours before the regular workday, specifically to vibe code without the pressure of their normal responsibilities.

As Natalia puts it: “Having that creative space is very counterintuitive to the way we usually work. How much of our time is really spent figuring out if there’s a new way to do things? Historically, you’re hired to do a specific set of functions until you get to the next level. For a company to say, ‘We think this is all changing and we trust you to figure it out’ — it’s really revolutionary.”

Case Study: Bespoke AI for Private Equity

One of Every’s longest-running engagements is with a private equity firm where the internal champion — a partner named Jonathan — understood something critical: AI adoption is not primarily a technical challenge. It’s a people challenge.

The Task-Mapping Foundation

Jonathan sat down with every investor at his firm and mapped out every single task they perform — from research to diligence to market mapping to portfolio management to just running their daily lives as investors. This produced an extraordinarily detailed view of what work actually looks like at the firm, broken down by team, since strategies vary significantly.

Every then went through that task map and highlighted where AI could deliver high leverage. The granularity made it possible to be precise about solutions rather than offering generic training.

”Savile Row” Prompt Tailoring

The firm had accumulated over a decade of investment thesis material in SharePoint — the intellectual property of how they think about opportunities. When diligencing a new company, investors needed to synthesize that entire body of knowledge against the new opportunity. Previously, this was an enormous manual task.

The solution involved three layers:

Data connection: Linking the AI to the firm’s SharePoint repositories and internal context
Custom prompts: Engineering prompts that reflect the specific way the investment committee analyzes opportunities — how numbers should appear, how figures are expressed, how the team thinks internally about their strategy
Output format: Generating draft investment memos that match the firm’s specific rubric and historical voice

Jonathan interviewed every investor and every team to capture these nuances. Without that degree of tailoring, the results would have been generic and useless.

Before: An analyst, associate, and principals spent 2–3 weeks drafting an investment memo before it went to the investment committee
After: A high-quality draft in approximately 30 minutes

The Broader Pattern

Dan identifies a general principle emerging from this case and others: connecting AI to your data sources is table stakes. The real value comes from encoding how your organization thinks — not just where the data lives, but how you define your metrics, how you reason about your domain, and where specifically to look for answers. This is what makes one company’s use of AI fundamentally different from another’s.

The Engineering Framework: Plan, Delegate, Assess, Compound

Every has identified a four-step cycle that works well for integrating AI into engineering organizations:

Plan: Define the scope and architecture of the problem
Delegate: Assign specific components to the AI
Assess: Review the generated code or output
Compound: Build upon successful outputs and learnings

The Missing Step Nobody Does

At one tech company Every is working with, the engineers were effective at delegating, assessing, and even compounding — but there was no planning phase. Without scaffolded plans, they could solve small issues but couldn’t address the large, systemic problems where AI delivers the biggest returns. They kept running into the same challenges because there was no architecture for the AI to work within.

Once they added proper planning to the workflow, the results changed dramatically. Natalia reports they’re consistently seeing engineers generate two weeks of effective work in a single afternoon when the plan-delegate-assess-compound framework is in place.

Technical Walkthrough: Claudia, the AI Project Manager

Claudia is an autonomous agent built in Claude Code (running on Opus 4.5) to manage Every’s consulting operations. The system lives in Every’s consulting GitHub repository.

Architecture

At the highest level, the system comprises five components:

claude.md (The Job Description): A persistent context file that Claudia reads every time it’s invoked. It defines the agent’s role, reporting lines, colleagues, client knowledge, and internal workflows. Natalia describes this as “the information a human project manager always knows — where they work, what their job is, what good looks like, who they report to, and who their colleagues are.”
Commands: Executable actions like Quality Check, Weekly Update, New Client Setup, and New Team Member Onboarding
Tasks: A dependency management system that breaks complex goals into phases and launches sub-agents to double- and triple-check work quality before returning results
Skills: General-purpose instruction files covering formatting standards, brand voice, and other conventions
MCPs (Model Context Protocol): Live data connectors to Gmail, Google Calendar, Google Drive, and meeting transcripts (Zoom/Granola)

Database Thinking for Unstructured Data

A key architectural unlock was applying relational database principles to text-based project management. Claudia uses ID conventions — Client ID, Person ID, Team ID — to map relationships between unstructured data (emails, call transcripts) and structured dashboards.

This means Claudia can answer questions like: Did this person attend this training session? Did that session deliver prompts or agents? How does this person relate to this team and this initiative? Without the ID system, connecting information across emails, calendars, meetings, and spreadsheets would be unreliable.

Operational Principles

The claude.md file encodes several principles that shape Claudia’s behavior:

Data accuracy is paramount — high fidelity is non-negotiable
Proactive, not reactive — don’t wait to be asked; anticipate necessary updates
Every interaction builds or erodes trust — consistency matters
Formulas over manual entry — when updating spreadsheets, write formulas rather than static numbers to maintain data integrity
When in doubt, escalate — ask questions rather than guessing

Natalia notes that the claude.md file is actually fairly concise. They don’t define what a project manager is — Claude is smart enough to know that. Instead, the file focuses on boundaries, conventions, and sharp edges specific to their operation.

Live Demo: New Client Setup

When Natalia runs the New Client Setup command for a client called Headway, the following sequence executes:

Skill loading: Claudia reads the “Client Work Skill” and handbook requirements
Information gathering: Four sub-agents launch simultaneously to scan:
- Gmail for contact details, scope, and correspondence
- Google Calendar for scheduled sessions
- Google Drive for existing assets
- Meeting transcripts for verbal context and commitments
Truth establishment: The agent synthesizes findings into a “foundational set of truths” about the engagement
Dashboard population: Client data flows into a structured Google Sheets dashboard

Dan pauses the demo to react: “You just launched four sub-agents to look through your Gmail, look through your calendar, look through your drive, look through your meetings to get contacts on the project, and then it’s going to go gather that information and put it in the right place into the spreadsheets that you use to run the business. That’s crazy.” Natalia’s response: “The only thing that’s crazier is that the alternative to Claudia doing this is me doing this.”

The setup process takes approximately 30 minutes — not instantaneous, but a fraction of the manual effort.

The Dashboard Output

The populated dashboard tracks:

Sessions: Total number of sessions to deliver, with status tracking
Deliverables: Training materials, workflows, curricula — each tagged and status-tracked
Open items: Automatically extracted from emails and meeting transcripts. If Natalia says “I’ll follow up on this” in a call, it appears as an open item with a priority level
People: Every contact has an ID, title, and team ID, mapping how they relate to each other and move across initiatives
Team summaries: Headcount per team, session participation, upcoming engagements
Session logs: Session ID, participating team, attendees, topics covered, delivery method, facilitator, attendance count
Source materials: Tagged and status-tracked
Feedback: Accumulated across the engagement

As sessions are scheduled, they’re automatically populated. When completed, they’re automatically marked as such. The entire system that previously required 10–15 hours per week of manual project management now requires about one hour.

How Claudia Was Actually Built

Three Failures Before Success

Claudia was scrapped and rebuilt three times before finding the architecture that worked. Each failure taught something:

Too engineering-focused: The first version was obsessed with framework and data architecture but lacked understanding of what good project management actually requires
Too vague: Another version was essentially just a job description — it described the work but didn’t have the structural scaffolding to execute it
The blend that worked: The final version combined Natash’s technical knowledge of Claude Code infrastructure (tasks, sub-agents, MCPs, commands) with Natalia’s deep understanding of what “good” looks like in consulting project management

The Pairing Model

The development process paired Natalia (the domain expert and subject matter expert on project management) with Natash Agarwal (an applied AI engineer, formerly at Quora). Neither could have built it alone — the engineering expertise without domain knowledge produced technically sound but practically useless systems, while domain knowledge without engineering structure produced descriptions that couldn’t execute.

Dan identifies this as a pattern he sees working well for CEOs: pair with someone who knows the technology edge, then tackle an ambitious project that’s half learning and half building something genuinely valuable.

The Shift from Operator to Manager

Building an effective agent creates a fundamental role change. Natalia’s work shifted from doing (tabulating data, manually updating sheets, searching through emails) to managing.

Ongoing Relationship Building

Just like a human employee, Claudia requires ongoing investment. When the agent makes a mistake or lacks sufficient information, Natalia goes back and updates the context files — commands, skills, or the claude.md — so the error doesn’t recur. “This is the same way you would build a relationship with any new staff member. You’re investing in that relationship as something you can rely on to get good work done.”

Reframing the Fear

Dan raises the common anxiety: if an agent can do a good portion of a job, what happens to the person? Every addresses this directly. During Think Week, they held a “Promote Yourself Day” where the premise was: figure out how to promote yourself so you’re not doing your IC job anymore — you’re one level above.

The framing makes it intuitive: if you hired a human project manager, you wouldn’t expect to lose your job. You’d expect to manage the project manager and take on more strategic work. The same logic applies to an AI agent.

Where the Time Goes

For Natalia, every hour not spent tabulating information is spent with clients. “My favorite thing about any work I’ve ever done has been working with people. Any hour I’m not in an Excel sheet, I’m spending with the people I get to work with. That is so much more fun and so much more valuable.”

Every’s consulting practice is available at every.to/consulting. The team includes Natalia (Head of Consulting), Natash Agarwal (Applied AI Engineer, formerly Quora), and Brooker Belor (Financial Practice Lead, formerly Perplexity), among others.

Marq AI Wiki

Explorer

How We Built 'Claudie,' Our AI Project Manager (Full Walkthrough)