Advice needed for adopting multi-agent workflow to Citadel

[raphaeld-zg]

Hello! Until now, before using Citadel, I've been using a three-pronged LLM collab setup at my day job. This human <-> LM <-> LLM process is slow but it delivers highly robust work, such as a recent bundle of PowerShell and Google App Script migration scripts that worked almost perfectly in my initial testing.

Each phase of my process in a nutshell:

1. Phase one involves ideation and planning. ChatGPT Pro Extended acts as a researcher, planner, and devil's advocate. We have a back-and-forth conversation to flesh out ideas and challenge our assumptions until we reach a point I deem good enough for ChatGPT to draft a high-level implementation plan.
2. Phase two is the context handoff. The high-level implementation plan is placed into an "agent-exchanges" folder within the local repository. The instructions that teach all involved LLMs how to read and create these communication reports are weaved not just into that folder as a readme, but throughout other files within the repo.
3. Phase three is hands-on execution. A local repo-aware "executing" LLM, like Claude Code or Codex, takes the plan from the folder, drafts its own step-by-step blueprint, and executes the work in the repository.
4. Phase four is the review and iteration loop. That executing LLM writes a summary of its work back into the agent exchanges folder. The updated repository is zipped and given back to ChatGPT Pro for review, allowing me to then repeat the process.

Now that I'll be using Citadel, what's the best way to go forward with this sort of setup? Is it as simple as using Claude to modify the necessary Citadel files so that any LLM working with that repo's aware of this sort of collaboration, at least at a high level?

Thank you! 🙏

[SethGammon]

Hey Zg!

You’re already thinking about this in the right shape. Citadel is built for exactly this kind of human-led, multi-agent workflow. The main difference is that it gives the handoff, memory, routing, and verification a repo-native structure instead of leaving you to maintain a custom agent-exchanges folder.

You can bring your existing ChatGPT planning phase into Citadel, or you can run the whole loop inside Citadel. If you want the Citadel-only version, I’d map your current phases like this:

1. Ideation and planning becomes .planning/intake/ plus a campaign brief.
2. Devil’s advocate / critique becomes an explicit review pass before implementation.
3. Execution runs through /do, a campaign phase, or Fleet depending on scope.
4. Review becomes the campaign artifact, git diff, commands run, verification result, unresolved risks, and next recommended action.

Practically, I’d start project-local:

• install/configure Citadel in the target repo
• run /do setup --express
• put repo conventions in AGENTS.md / CLAUDE.md
• use .planning/intake/ for raw requests and rough ideas
• use campaigns for work that needs to survive across sessions
• use .Codex/agent-context/ for Codex-specific context
• use Fleet only when the work can genuinely split into mostly independent lanes

I would not start by asking Claude/Codex to modify Citadel core files. Start by configuring the target repo. Only change Citadel itself if you discover a general workflow Citadel should support for every user, not just one repo.

The big shift is that agents should stop writing loose summaries into a custom exchange folder and start leaving structured repo evidence:

• files changed
• commands run
• verification result
• decisions made
• unresolved risks/questions
• next recommended action

So the short answer is: yes, you can keep your workflow shape and do it entirely inside Citadel. Use Citadel’s planning, campaign, and fleet structure as the shared memory layer, then let /do and the campaign artifacts carry the work from plan to implementation to review.

You can do your workflow with Citadel as the execution layer (and architecture) while still using something external like ChatGPT, or you can do it all in Citadel.

Feel free to come back with any questions or anything you'd like to discuss further!