I use a series of stop hook [0] scripts. For example, I have a script [1] that forces Claude to execute tests whenever code files are changed. The stop hook runs automatically and will force Claude to continue working until the script passes.
I also have a script that forces Claude to generate a summary of work [2] if it hadn't done so on it's own.
I don't think you understand what I'm saying. If, during a session, Claude makes any change to a source file, the stop hook script FORCES Claude to run the existing tests. There is literally no way Claude can get around running the tests because the prompt will not stop being processed until the stop hook script passes.
There is no contradiction. Stop hooks (as well as all the other hook types) are the only way to force Claude to work deterministically.
Hook scripts can be as simple or as complex as you like: you define the success criteria. For example, if Claude just added a new feature but didn't create a test for it, then a stop hook would prevent Claude from stopping until the test was written.
stop hook forces the claude to run the existing tests. Fine
if claude added a new feature but didn't created a test for it. It will wait for it to perform that action. Fine.
We can adjust the complexitiy of hook scripts. That's ok.
Tell me if I'm wrong, I am understanding it more like a compiler like if the syntax is ok, just pass. Similarly here, if the test were ran, it will look for a marker file in /tmp, it found it, and pass.
I did not understood the part of tests. Maybe my question is more clear now.
This is one of the simplest and most fundamental ways to manage Claude’s behavior. I genuinely can’t make it clearer. The issue you’re describing seems to center on situations where Claude becomes “stubborn” and doesn’t follow instructions.
When that happens, the solution is straightforward: create a hook script that explicitly enforces the behavior you want. By doing so, you remove ambiguity and leave Claude no option but to comply.
If you can share a specific case where Claude isn’t following your prompts, I can help you craft an appropriate hook to correct it.
Haven't really experienced it besides in long contexts. Usually keep it concise. Something like:
We're building feature X.
- You might need `a`, `b`, `c`. (any libraries, url documentation etc)
- The requirements are:
- x
- y
- z
Any negatives (prevent going down a path) would go into requirement.
This would be a prompt to sisyphus if small or prometheus if big. (Using opencode + oh-my-openagent).
If I believe the agent won't understand what it is supposed to do or if there are multiple solutions of which only some are allowed I add "DO NOT make or edit any business rules before asking me".
> We're building feature X.
- You might need `a`, `b`, `c`. (any libraries, url documentation etc)
- The requirements are:
- x
- y
- z
So, you are saying we have to start every conversation for an existing project with this context? pointing to the particular feature detail we are targeting?
We have a bug in our handler for the distributed transaction which updates the payment amount. It allows modifications when it should not.
The requirements are:
- We should reject regular changes outside of intent state
- We should allow upsell when state is accepted and upsell modifications is activate in the config
- You do NOT have to consider shipping during upsell, it is out of scope and is already required not to change as a result of upsell elsewhere.
We don't use any CLAUDE.md files or similar. Skills for things that it gets wrong and the rest it can handle by reading the repos themselves.
Underlying plugin prompts makes sure it completes the task and adds tests etc.
Now it is convincing. But the same approach a week ago frustrated me. The problem was some guy in office wrote the logs, and it was somehow logging success even though it was going in the catch block due to an error. But the claude was convinced that the logs were right and even though i was telling it not to trust the logs and listing the exact thing to fix, it did the opposite and in the end i had to do it manually along with fixing the logger. Nevermind, your approach is reliable and will surely work when the logs are fine. Thanks
I use a series of stop hook [0] scripts. For example, I have a script [1] that forces Claude to execute tests whenever code files are changed. The stop hook runs automatically and will force Claude to continue working until the script passes.
I also have a script that forces Claude to generate a summary of work [2] if it hadn't done so on it's own.
[0]: https://code.claude.com/docs/en/hooks
[1]: https://gist.github.com/Looking4OffSwitch/c3d5848935fec5ac3b...
[2]: https://gist.github.com/Looking4OffSwitch/3b13b65e40284be899...
Looks an interesting approach.
So, the tests being written are based on what? on user input or to test the changes it made according to claude's analogy?
If it is the latter, it can be the same problem of forcing its analogy. In this case by validating it
> So, the tests being written are based on what?
I don't think you understand what I'm saying. If, during a session, Claude makes any change to a source file, the stop hook script FORCES Claude to run the existing tests. There is literally no way Claude can get around running the tests because the prompt will not stop being processed until the stop hook script passes.
There is no contradiction. Stop hooks (as well as all the other hook types) are the only way to force Claude to work deterministically.
Hook scripts can be as simple or as complex as you like: you define the success criteria. For example, if Claude just added a new feature but didn't create a test for it, then a stop hook would prevent Claude from stopping until the test was written.
Mate I still didn't understand that.
stop hook forces the claude to run the existing tests. Fine if claude added a new feature but didn't created a test for it. It will wait for it to perform that action. Fine. We can adjust the complexitiy of hook scripts. That's ok.
Tell me if I'm wrong, I am understanding it more like a compiler like if the syntax is ok, just pass. Similarly here, if the test were ran, it will look for a marker file in /tmp, it found it, and pass.
I did not understood the part of tests. Maybe my question is more clear now.
This is one of the simplest and most fundamental ways to manage Claude’s behavior. I genuinely can’t make it clearer. The issue you’re describing seems to center on situations where Claude becomes “stubborn” and doesn’t follow instructions.
When that happens, the solution is straightforward: create a hook script that explicitly enforces the behavior you want. By doing so, you remove ambiguity and leave Claude no option but to comply.
If you can share a specific case where Claude isn’t following your prompts, I can help you craft an appropriate hook to correct it.
Haven't really experienced it besides in long contexts. Usually keep it concise. Something like:
Any negatives (prevent going down a path) would go into requirement.This would be a prompt to sisyphus if small or prometheus if big. (Using opencode + oh-my-openagent).
If I believe the agent won't understand what it is supposed to do or if there are multiple solutions of which only some are allowed I add "DO NOT make or edit any business rules before asking me".
> We're building feature X. - You might need `a`, `b`, `c`. (any libraries, url documentation etc) - The requirements are: - x - y - z
So, you are saying we have to start every conversation for an existing project with this context? pointing to the particular feature detail we are targeting?
No this would be something targeted such as:
We don't use any CLAUDE.md files or similar. Skills for things that it gets wrong and the rest it can handle by reading the repos themselves.Underlying plugin prompts makes sure it completes the task and adds tests etc.
Now it is convincing. But the same approach a week ago frustrated me. The problem was some guy in office wrote the logs, and it was somehow logging success even though it was going in the catch block due to an error. But the claude was convinced that the logs were right and even though i was telling it not to trust the logs and listing the exact thing to fix, it did the opposite and in the end i had to do it manually along with fixing the logger. Nevermind, your approach is reliable and will surely work when the logs are fine. Thanks