# Let architects build the AI setup once and roll it out to users and groups

**URL:** <https://community.fibery.io/t/let-architects-build-the-ai-setup-once-and-roll-it-out-to-users-and-groups/11651>\
**Category:** Ideas & Features\
**Created:** [November 14, 2026, 9:48pm UTC](https://community.fibery.io/t/let-architects-build-the-ai-setup-once-and-roll-it-out-to-users-and-groups/11651 "2026-11-14T21:48:08Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![tstirm](https://avatars.discourse-cdn.com/v4/letter/t/df705f/32.png) [@tstirm](https://community.fibery.io/u/tstirm)\
**Post date:** [November 14, 2026, 9:48pm UTC](https://community.fibery.io/t/let-architects-build-the-ai-setup-once-and-roll-it-out-to-users-and-groups/11651/1 "2026-11-14T21:48:08Z")

</div>

This started as a +1 on [Hide Fibery AI from members or allow only for specific user groups](https://community.fibery.io/t/hide-fibery-ai-from-members-or-allow-only-for-specific-user-groups/11447). But I think per-user credits are just the first piece of a bigger ask: a better AI harness. A harness is the setup around the AI model (its instructions, knowledge and skills) that turns a general chatbot into one that knows how you and your team work.

**The problem**

I use Claude heavily to build and run my Fibery workspace. Over time I’ve built a setup that uses every part of an AI harness:

- default context that always loads: who I am, how I want it to work, and every Fibery MCP gotcha and shortcut I’ve learned

- curated knowledge it pulls from

- a library of skills for repeatable tasks, including standard steps to start and close a working session

- working memory between chats, saved back into Fibery so the next chat picks up where the last one stopped

- a feedback loop: when something goes wrong, the fix goes back into the setup so it doesn’t happen again

It makes me a lot more effective. Where I can’t bring this setup with me, I notice the difference.

No normal user is going to build something like this. Even with the AI doing most of the upkeep, it takes real effort to design, and I only do it because the payoff is worth it to me. Today, sharing it with a team means either handing people zip files to install or standing up Claude Team/Enterprise. That’s a second platform, with its own users, groups and content to keep in step with Fibery.

**The ask**

Let architects build the AI setup once, in Fibery, and roll it out to users and groups.

Users should be able to configure it too. Architects set sensible defaults, and each user adds their own context, knowledge and skills on top, without breaking anyone else’s setup.

Fibery is the natural home for this. It already has the groups and permissions, and more importantly the data model. A skill or a piece of knowledge can be an entity related to the Spaces, databases and groups it applies to, with a workflow to review it before it goes live. Notion already has part of this: skills as pages, and default workspace skills per group. Fibery’s data model lets it go further.

I’m not asking Fibery to become an AI product. I’m asking it to let AI use what Fibery already does best: the data model, relations and workflows. That’s what would make a Fibery harness better than one built on pages.

**What it would include**

1. **Default context per group.** Always-loaded instructions: who the group is, how the AI should work, and the workspace’s conventions and gotchas. It’s small and always included in full, so it never depends on search finding it.

2. **Knowledge per group.** Architects point the AI at curated entities and docs, such as methods, procedures and reference material. Semantic search can look in the group’s or owner’s content first, so answers aren’t diluted by everything that happens to be visible.

3. **Skills per group.** Skills stored as entities, each with a name, a description, the instructions themselves, and relations to the groups, Spaces and databases it applies to. A review workflow controls when it goes live. Updates reach everyone automatically, with no zip files. Only names and descriptions load up front, and the full skill loads when it’s needed (the open Agent Skills approach), which keeps credit use down.

4. **Credits per user and group.** Turn Fibery AI on or off per user or group, set caps per billing period, see usage by user and group, and give users a clear message when they hit a cap. This is what makes rolling the rest out to everyone safe. Once Fibery AI is everyone’s default tool, one shared credit pool won’t hold up.

5. **Open format.** Import and export standard `SKILL.md` files, and serve skills and context over Fibery’s MCP. Fibery becomes where an organization keeps its AI setup, whichever AI tool reads it.

**Why it’s worth it for Fibery**

- Today, power users who want a setup like this build it in Claude, ChatGPT or another AI tool connected to Fibery over MCP, and Fibery AI gets skipped. If architects could build that setup inside Fibery, Fibery AI would become the easiest choice for every member, not just the people who build the workspace.

- The setup becomes something an organization invests in inside Fibery. That’s hard to walk away from.

- Caps make it safe to give everyone AI access, so more people use it. Spend an organization can control is spend it’s willing to raise.

- Notion has shown there’s demand. Fibery’s data model is a real edge in meeting it.