I Ran My Firm's Discovery on a Pile of Claude Skills. Here Is What I Built Instead

If you run a small consulting firm and your clients keep asking you for AI advice faster than you can credibly give it, you have probably done what I did. You started stacking Claude skills. One for reading client documents. One for drafting interview questions. One for turning a messy transcript into a gap analysis. A course PDF you bought in a panic. A Notion page only you understand. From the outside it looks like a process. From the inside you know it is a pile.

I lived in that pile for the better part of a year while building a discovery process I could actually hand to other people. This post is the origin story of why the pile fails and what I built once I finally understood the real problem. I am not figuring out my thesis here. I proved it the hard way. This is the construction log.

The pile worked, right up until it didn't

Here is the honest version. The pile was good. I had maybe thirty Claude skills doing real work. I could open a client's document dump, run my extraction skill, run my questioning skill, paste the output into my synthesis skill, and produce a discovery that genuinely impressed people. Clients could not tell my process from a Big Four method. That is the trap. It works well enough that you keep adding to it instead of questioning it.

Then a second engagement landed on the same week as the first, and the whole thing came apart.

Not because the skills broke. Because I was the runtime. The sequencing lived in my head. The judgment about which skill to run when, what "good" output looked like, which step to skip for a thin document set, none of that was written down anywhere. It was muscle memory. And muscle memory does not scale to two clients at once, let alone to an associate.

This is the founder bottleneck, and if you run a 3-to-25 person firm it is probably your real constraint even if you are blaming time or pipeline. The method is in your head. You cannot hand it off. Every engagement routes through you because you are the only person who knows the choreography. I have heard this exact story again and again across hundreds of conversations with consultants who are good at the work and stuck because the work cannot leave their hands.

Why "get better at AI" was the wrong fix

For a long time my answer to the bottleneck was to get better at AI personally. More skills. A new course. The latest model. I told myself that once I was proficient enough, I would be able to advise clients credibly and the firm would follow.

That belief is seductive and it is a treadmill. The technology does not hold still long enough for personal proficiency to become an asset. The day you feel caught up, the ground moves. To put a number on how fast the ground moves: general-purpose AI obligations under the EU AI Act entered into application on 2 August 2025, with the Commission's enforcement powers and fines arriving a year later. That is one regulation on one timeline. Now add a new frontier model every few weeks. A method you froze into a course PDF last quarter is already drifting.

The deeper problem is that personal proficiency is the wrong unit entirely. Your clients are not buying your knowledge of last month's model. They are buying a rigorous process they can trust. And the data backs the gap between knowing and delivering. MIT's NANDA initiative found that roughly 95% of enterprise generative AI pilots never make it past the pilot phase, and the cause was organizational, not technical. The firms that win are not the ones who learned the most AI. They are the ones who built a process that holds.

That reframe is the whole thing. Stop chasing the edge. Stand on infrastructure that holds it for you. Once I believed that, the pile of skills stopped looking like progress and started looking like the symptom.

What I built instead: a system, not a smarter pile

The fix was not a better skill. It was turning the pile into rails.

A tool does one task. A system sequences the tasks, enforces the order, captures the output in a consistent format every time, and lets someone other than me run it. That last clause is the whole game. If a junior person cannot run the front half of discovery without me in the room, I do not have a system. I have a hobby that pays well.

So I rebuilt my discovery as one connected workflow instead of a drawer of disconnected tools. Concretely, that meant:

• One front door for intake. Instead of me deciding what to ask for, the system asks for it. Specific documents tied to specific workflows, not "send me whatever you have." Standardizing intake with role-specific questionnaires was the single highest-leverage change, because it moved the first decision off my plate.
• Document analysis that runs the same way every time. The extraction step used to be thirty minutes of me deciding how to prompt. Now it is a fixed step with a fixed output shape. The point of structured document analysis is not speed. It is repeatability. Same input quality, same output quality, regardless of who clicks run.
• Discovery questions generated from the client's actual data, not template prompts. This is where the firm proves it did its homework, and it is also where most piles leak, because the questioning skill in your drawer does not know what your extraction skill found unless you personally carry that context between them.
• A synthesis layer the lead consultant enters at, after the associate has run everything upstream. The judgment stays human. The choreography stops being human.

The result is the thing I could not buy: a discovery that runs without me. An associate carries intake, extraction, and the first call. I enter at synthesis, where the strategic read actually lives. If you want the step-by-step version of that handoff, I wrote up exactly how my team runs it.

The part the pile could never give me

Two things changed that I did not expect.

First, the system made me sharper, not lazier. I assumed that automating the front half would dull my edge. The opposite happened. When extraction and questioning are consistent, the only thing left for me to do is the judgment, and I do more of it, on more clients, with more attention. Proficiency turned out to be a byproduct of running good rails, not a prerequisite I had to grind out first.

Second, the system does not go stale on its own the way my pile did. A pile of skills is frozen the moment you stop maintaining each one, and you will not maintain thirty things. Infrastructure that ingests the latest models and rules continuously means my firm's edge compounds instead of decaying. I am not personally responsible for keeping up anymore. The rails keep up. I keep advising.

That is the difference between a tool and a platform you build your practice on. A tool is something you operate. A platform is something your firm operates through, including the people who are not you.

Where this points next, for your team

I built this to solve the founder bottleneck, my problem. But the moment you have associates, the same disease shows up one level out. Your people each run discovery their own way. One associate's intake looks nothing like another's. The deliverables drift. The methodology you were so proud of fragments across the bench, and now you have a consistency problem on top of a capacity problem.

That is the next thing the system solves, and honestly the bigger prize. When discovery runs on shared rails, everyone runs it the same way, and your method stops being a thing that lives in any one head. It becomes infrastructure your whole team stands on. If you run a firm with more than one person in it, productizing discovery so the team runs it consistently is the work that turns a talented founder into a firm.

This is the part I am building in the open now. The thesis is closed. I lived the sprawl, I heard the same story across hundreds of consultant conversations, and I spent a year turning the pile into a system. What I am building, Audity, is the version of that system you do not have to assemble yourself.

Audity is a white-label AI readiness assessment platform for consulting firms. It lets a boutique firm run a repeatable AI readiness assessment under its own brand and turn the findings into a client-ready deliverable, so the diagnostic itself produces the qualified pipeline instead of one more PDF the founder has to assemble by hand. The method leaves your head, your associates run discovery the same way every time, and the platform keeps ingesting the latest models and rules so your firm's edge compounds.

The bottom line

Running your firm's discovery on a pile of Claude skills is not a mistake. It is a stage. Almost everyone good starts there, because the pile genuinely works until you try to scale it or hand it off. The mistake is staying there, adding skill number thirty-one, and calling fragmentation a process.

The move is to stop collecting tools and start building rails. Sequence the work. Fix the output shape. Make it runnable by someone who is not you. Then let the infrastructure stay current so your firm's edge compounds instead of decaying. That is how the method leaves your head, your associates run discovery the same way every time, and you finally get to spend your hours on the judgment clients actually pay for.

I had a pile. I built a system. The system is the practice.

Sources

• AI Act overview and implementation timeline, European Commission, Shaping Europe's digital future
• MIT report: 95% of generative AI pilots at companies are failing, Fortune (Aug 2025)

---

Built for boutique consulting firms

Audity is the operating system for boutique firms productizing their discovery process and running premium engagements at speed. If you run a team, your lead consultant is the bottleneck, and you want associates closing engagements without losing methodology integrity, this is built for you.

See how Audity works for your team →