llm-tooling (former ChatGPT-Primer)

This repo contains claude code tooling and my old ChatGPT Primer.

How it Works

The claude code harness tries to give claude hints and rules to avoid the most common llm failure modes. It works by adding pre and after hooks to the answers, where scripts react to the in and outputs of human and machine, kindly reminding claude to BE AN ADULT.

Hints

If you spot an error: give additional context which avoids the error and re-prompt again.
DO NOT attempt to "fix" the error within the same thread! This will only lead to context poisoning.

Guiding principles

The Machine will try to keep to the facts and generally be as grounded as possible. If the Machine steers off, it will get reprimanded. If the user types stuff which leads to known llm failure modes, the harness injects preemptive reminders. The Machine will forget. It's on the harness to work around this.

Background

Rules collected over my work with cc and constantly getting reminded that llms like to acknowledge things while changing absolutely nothing.

As a plugin, which installs the hooks and the skills together:

/plugin marketplace add Kalabint/llm-tooling

Or manually, if you want them in your own ~/.claude:

git clone [email protected]:Kalabint/llm-tooling.git

Copy .claude/hooks/ into ~/.claude/hooks/ and skills/ into ~/.claude/skills/, then merge .claude/settings.hooks.json into your ~/.claude/settings.json.

Either way, copy .claude/hooks/local.env.example to ~/.claude/hooks/local.env and fill in what applies. Every key is optional and an unset one makes that check inert, so the harness runs unconfigured.

Example

User:  did the loop approach work better?
Hook:  VERDICT ASKED -> name the quantity asked about vs the proxy you can
       measure, and try to refute your own strongest number first.
       [skill: .claude/skills/discriminator/]

Model: "my earlier count was wrong" with no re-run
Hook:  BLOCK -> re-run it now and give the corrected number. Re-wording does
       not clear this.

Repository Contents

• .claude/skills/: the skills. Start with discriminator, no-invented-hazards, check-dont-assume.
• .claude/hooks/: the hooks and the shared pattern file, listed with their lifecycle events in its README, which is generated. Needs bash and jq.
• scripts/: pii-sweep.sh refuses to publish identifying data and runs from the git hooks; corpus-stats.sh measures how often each guard would fire against your own transcripts.
• universal_task_system_instructions.txt: the ChatGPT primer — give it a query, it lists topics, selects an overqualified role, and answers from that perspective.
• chatgpt_custom_instructions.md: custom-instructions form, both fields.