We built a complete AI workspace for a professional services firm — an IT services consultancy working across ITAM, SAM, and ITSM with a team of 30+ people.
The workspace covers 18 workflow skills across three core domains: tender evaluation, proposal generation, and marketing content production. This post is about what we actually learned doing it — the architecture decisions, the mistakes, and the things that mattered more than we expected.
Lesson 1: The knowledge base is the product, not the AI
This was the biggest insight and it took longer than it should have to fully internalise. The skills themselves are almost trivial once the knowledge base is solid. A well-structured prompt pointing at accurate, comprehensive company knowledge produces good output reliably.
If the knowledge base is weak, no amount of prompt engineering helps. You can rewrite a prompt fifteen times, but if the underlying facts are wrong or missing, the output will be wrong or generic. We stopped debugging prompts and started debugging knowledge. That was the turning point.
Lesson 2: Skills should be small and chainable
Early on we tried building monolithic skills — one skill that takes a tender document and produces a complete proposal. It was brittle, hard to debug, and the quality was inconsistent.
The version that works: each skill does one thing. Proposal generation is actually four sub-skills: calculate pricing, generate statement of work, write technical specification, build the presentation deck. Each one can be run independently. Each one can be checked independently.
When something goes wrong, you know exactly where. When quality drops in one area, you fix one skill without touching the rest.
Lesson 3: A quality gate changes everything
We added a shared quality checklist that runs on every output. Brand voice compliance. Factual accuracy against the knowledge base. Formatting standards. Commercial rule adherence. Completeness checks.
This was the single highest-impact decision in the entire build. Before the quality gate, outputs were inconsistent — some great, some off. After it, the floor rose dramatically. Bad outputs still happen, but they are caught before they reach anyone.
Lesson 4: Document generation is harder than it looks
Producing branded .docx and .pptx files from AI output is genuinely difficult. This is not a glamorous insight, but it is an honest one.
Template compliance. Formatting consistency. Table handling. Image placement. Style inheritance. Header and footer logic. These problems are tedious and time-consuming.
This is where roughly 40% of the total build time went. If you are scoping a similar project, do not underestimate this. The AI can generate excellent content in minutes. Getting that content into a properly formatted branded document takes real engineering work.
Lesson 5: The CLAUDE.md file is the control centre
One file defines the system identity, skill routing, and core rules. Every session starts by reading this file. It tells the AI who it is, what it can do, what rules it must follow, and how to route different requests.
This is what makes the workspace feel coherent rather than like a collection of disconnected prompts. Without it, each skill operates in isolation. With it, there is a shared understanding that carries across every interaction.
Lesson 6: Projects mirror for team access
Not everyone uses Claude Code. Building a parallel Projects setup in Claude.ai for colleagues without CLI access kept the whole team working from the same knowledge foundation.
Same knowledge base. Same quality standards. Different interface. This was important for adoption. If only the technical team can use the system, the business value is limited.
Lesson 7: Iteration beats planning
We rebuilt the tender evaluation scoring model three times based on user feedback. The first version was overengineered — too many criteria, weighted scoring that nobody understood, outputs that were longer than the original tender.
The final version is simpler and more accurate. It scores what actually matters for bid/no-bid decisions and presents results in a format the team can act on in minutes.
We could not have designed the final version on day one. We needed to see it fail in practice to understand what good looked like.
The numbers
- 15 knowledge base files
- 18 workflow skills
- 3 core domains: tender evaluation, proposal generation, marketing content
- 13+ marketing content formats (blog posts, case studies, social media, newsletters, and more)
- Quality gate on every single output
Advice if you are building something similar
Start with the knowledge base. Seriously. Do not build a single skill until your company knowledge is structured, accurate, and comprehensive.
Build small. One skill, one job. Chain them together for complex workflows.
Add a quality gate early. It is the cheapest way to raise output quality across the board.
Budget real time for document generation. If your outputs need to be branded documents, that is a significant workstream on its own.
And iterate. Your first version will not be your best. Build it, use it, fix it, repeat.