Improvement for claude.md

[umag]

I suggest to add line that forces claude to build list and get methods for swamp models firstly and then all the crud - that helps in having reliable data first and then using it in mutating workflows

[github-actions]

Hi @umag, thanks for opening this issue!

We appreciate you taking the time to report it. A maintainer will review it as soon as possible.

In the meantime, please make sure you've included all relevant details (steps to reproduce, expected vs actual behaviour, swamp version, etc.) to help us triage it quickly.

[johnrwatson]

Hey @umag,

One of the troubles with the suggestion above is that list and get methods might be totally inappropriate for some models. For example, as a user if you required and authored a model that was all about deleting AWS accounts + the model took just the account ID, neither GET or LIST are required for the model to work/will cause token churn and would waste time unnecessarily.

My recommendation here would be when prompting the agent to build the models, use plan mode and specifically call out the methods you want, that will hugely increase the probability for accurate and fast results.

Thanks,
John

[umag]

That's a valid objection, alas that would be a model in vacuum,
for example one case from discord chat where claude hallucinated public ip for a hetzner machine, that would be prevented if first step would be read model and only after that do the action.

And even in case where you want only to delete AWS accounts, you would need somehow get account id, (suppose thats an external knowledge), run deletion, and verify that it is absent - having read model automates first and last steps and puts that information explicitly into the context and you can work with that further

[stack72]

Hey @umag, thanks for the suggestion!

After thinking about this, we don't think adding this to CLAUDE.md is the right approach. As John mentioned, not all models need list and get methods — the methods a model exposes should be driven by the domain it represents, not by a blanket rule.

For example, a Hetzner server model naturally needs list and get because reading state is core to that domain. But a model focused purely on a destructive action like deleting an account doesn't need them, and forcing their creation would add unnecessary complexity and token churn.

Plan mode already handles this well — when you use plan mode and describe what you're trying to automate, Claude will analyze the domain and choose appropriate methods. If you need list and get, just call that out in your prompt and it'll include them.

We're going to close this one out, but appreciate you raising it!

Paul