I've faced the same daily grind of repeating codebase context and prompts to AI every time. Shared CLAUDE.md files fix that beautifully, and you could do the same for security agents centralizing team rules so everyone stays safe by default.Also it decreases the load on the agent to re load everything again and again.
I have saved Hannah's Claude.md Guide in the past for it offered so much value and learnings. It's so awesome to see in this post how you guys have identified communication gaps and lack of co-ordination in an organization as the real bottleneck and how, as a collective, one can take a leaf of Claude's booklet to create a shared context file. 🌟
The compounding correction point is underrated. Most teams treat context files as static documentation rather than living systems -- so fixes stay local instead of propagating. What I'd add from solo use: even without a team, the same dynamic applies across sessions. The file grows smarter over time only if you structure it so updates in one section don't create noise in another.
The 6,000 hours math hit me, I never quantified it but I feel it every time I onboard someone new to my projects. I've been living in CLAUDE.md files for a year now. My content-engine repo has one that tracks lessons learned, naming conventions, even which Gumroad handle to use (I kept getting it wrong 😅).
The part about corrections compounding is real, every time Claude makes a mistake and I add it to CLAUDE.md, it never makes that mistake again in that project. One thing I've found though: the files drift if nobody owns them. Mine gets updated multiple times a week but only because I treat it like living documentation, not a setup-and-forget thing.
Curious how you'd handle that in a team setting where 10 people are all adding corrections does it get noisy fast?
@Jenny Ouyang Great question! While it has a learning curve, GitHub is the best tool for collaboration on shared context files. However, you can set up lighter weight processes with Google Drive / Notion / Dropbox. What really matters is establishing clear team processes and ways of working.
Moving AI into the shared drive/folder structure makes a lot of sense for the systems that were designed to mirror the concept and core company infrastructure. It’s the difference between everyone having their own calculator and everyone using the same real-time ledger. I love the 're-explanation tax' metaphor, great collab Joel and Hannah!
I’ve seen people who have outgrown roles be invited to stay at the company because they have “institutional knowledge”. A shared context sheet fed into an AI chatbot starts to eliminate the need for that…
This is a great point. Shared context files (when used effectively) can capture a lot of what used to only live in someone's head - the decisions that were made, the reasons behind them, and how things work in a particular organization. But, there will always be context that can't be written down. This is the context that informs judgment calls, creative leaps, and the ability to see connections that aren't obvious yet. That's where the value will shift - from "I'm the only one who knows how this works" to "I'm the one who sees what to do next."
Great article Hannah and Joel!
I've faced the same daily grind of repeating codebase context and prompts to AI every time. Shared CLAUDE.md files fix that beautifully, and you could do the same for security agents centralizing team rules so everyone stays safe by default.Also it decreases the load on the agent to re load everything again and again.
Yes! It adds simplicity AND decreases the load, so true!
A1 post, Hannah and Joel!
I have saved Hannah's Claude.md Guide in the past for it offered so much value and learnings. It's so awesome to see in this post how you guys have identified communication gaps and lack of co-ordination in an organization as the real bottleneck and how, as a collective, one can take a leaf of Claude's booklet to create a shared context file. 🌟
The organizations that align on a shared context strategy will reap the rewards!
Without context human too is dumb.
The compounding correction point is underrated. Most teams treat context files as static documentation rather than living systems -- so fixes stay local instead of propagating. What I'd add from solo use: even without a team, the same dynamic applies across sessions. The file grows smarter over time only if you structure it so updates in one section don't create noise in another.
After a year of daily Claude Code sessions I ended up splitting by concern rather than keeping one monolithic file. Makes targeted updates much cleaner. Breakdown of that architecture: https://thoughts.jock.pl/p/how-i-structure-claude-md-after-1000-sessions
That’s a very interesting idea to split them!
Joel, thanks for introducing Hannah!
The 6,000 hours math hit me, I never quantified it but I feel it every time I onboard someone new to my projects. I've been living in CLAUDE.md files for a year now. My content-engine repo has one that tracks lessons learned, naming conventions, even which Gumroad handle to use (I kept getting it wrong 😅).
The part about corrections compounding is real, every time Claude makes a mistake and I add it to CLAUDE.md, it never makes that mistake again in that project. One thing I've found though: the files drift if nobody owns them. Mine gets updated multiple times a week but only because I treat it like living documentation, not a setup-and-forget thing.
Curious how you'd handle that in a team setting where 10 people are all adding corrections does it get noisy fast?
@Jenny Ouyang Great question! While it has a learning curve, GitHub is the best tool for collaboration on shared context files. However, you can set up lighter weight processes with Google Drive / Notion / Dropbox. What really matters is establishing clear team processes and ways of working.
Tracking lessons learned is such a good idea!!!
Great question on handling that for a team, leaving that one for @Hannah Stulberg
The folder structure is crazy impressive.
This is post is an essential for leaders implementing AI. Well done!🩷🦩
Thank you! Truly a must for AI-forward organizations.
This should definitely be a practice that business teams start adopting in 2026. It keeps everyone on the same page and saves a ton of time.
It does!! And diminishes risk of people going rogue
Thanks for bringing Hannah to our attention... a bit of gold right there
Yes! 🙌
10x setup that's a must-do
It is!!
Moving AI into the shared drive/folder structure makes a lot of sense for the systems that were designed to mirror the concept and core company infrastructure. It’s the difference between everyone having their own calculator and everyone using the same real-time ledger. I love the 're-explanation tax' metaphor, great collab Joel and Hannah!
I thought you’d enjoy this one, @Anna!
Absolutely love the own calculator versus real-time ledger metaphor! Great way to describe the benefits of the shared context approach.
That’s a great idea! 💡
I’ve seen people who have outgrown roles be invited to stay at the company because they have “institutional knowledge”. A shared context sheet fed into an AI chatbot starts to eliminate the need for that…
Really interesting point, Anika!
This is a great point. Shared context files (when used effectively) can capture a lot of what used to only live in someone's head - the decisions that were made, the reasons behind them, and how things work in a particular organization. But, there will always be context that can't be written down. This is the context that informs judgment calls, creative leaps, and the ability to see connections that aren't obvious yet. That's where the value will shift - from "I'm the only one who knows how this works" to "I'm the one who sees what to do next."