1. openclaw like - using the LLM endpoint on subscription billing, different prompts than claude code

2. using claude cli with -p, in headless mode

The second runs through their code and prompts, just calls claude in non-interactive mode for subtasks. I feel especially put off by restricting the second kind. I need it to run judge agents to review plans and code.

- pre file write -> block editing code files without a task and plan of work

- post tool use -> show next open checkbox in the task to the agent, like an instruction pointer

- post user message -> log all user messages for periodic review of intent alignment

These 3 hooks + plain md files make my claude harness.

Work is divided into individual tasks. I could have used Plan Mode or TodoWriter tool to implement tasks - all agents have them nowadays. But instead I chose to plan in task.md files because they can be edited iteratively, start as a user request, develop into a plan with checkbox-able steps, the plan is reviewed by judge agent (in yolo mode, and fresh context), then worker agent solves gates. The gates enforce a workflow of testing soon, testing extensively. There is another implementation judge again in yolo mode. And at the end we update the memory/bootstrap document.

Task files go into the git repo. I also log all user messages and implement intent validation with the judge agents. The judges validate intent along the chain "chat -> task -> plan -> code -> tests". Nothing is lost, the project remembers and understands its history. In fact I like to run retrospective tasks where a task.md 'eats' previous tasks and produces a general project perspective not visible locally.

In my system everything is a md file, logged and versioned on git. You have no issue extracting your memories, in fact I made reflection on past work a primitive operation of this harness. I am using it for coding primarily, but it is just as good for deep research, literature reviews, organizing subject matter and tutoring me on topics, investment planning and orchestrating agent experiment loops like autoresearch. That is because the task.md is just a generic programming pipeline, gates are instructions in natural language, you can use it for any cognitive work. Longest task.md I ran was 700 steps, took hours to complete, but worked reliably.

https://github.com/horiacristescu/claude-playbook-plugin

I think this is a simplification, of course Bob relied on AI but they also used their own brain to think about the problem. Bob is not reducible to "a competent prompt engineer", if you think that just take any person who prompts unrelated to physics and ask them to do Bob's work.

In fact Bob might have a change to cover more mileage on the higher level of work while Alice does the same on the lower level. Which is better? It depends on how AI will evolve.

The article assumes the alternative to AI-assisted work is careful human work. I am not sure careful human work is all that good, or that it will scale well in the future. Better to rely on AI on top of careful human work.

My objection comes from remembering how senior devs review PRs ... "LGTM" .. it's pure vibes. If you are to seriously review a PR you have to run it, test it, check its edge cases, eval its performance - more work than making the PR itself. The entire history of software is littered with bugs that sailed through review because review is performative most of the time.

Anyone remember the verification crisis in science?

One interesting data point - I counted word count in my chat messages vs final code and they came out about 1:1, but in reality a programmer would type 10x the final code during development. From a different perspective I found I created 10x more projects since I relied on Claude and my harness than before. So it looks user intent is 10x more effective than manual coding now.

I also make individual tasks md files (task.md) which makes them capable of carrying intent, plan, but not just checkbox driven "- [ ]" gates, they get annotated with outcomes, and become a workbook after execution. The same task.md is seen twice by judge agents which run without extra context, the plan judge and the implementation judge.

I ran tests to see which component of my harness contributes the most and it came out that it is the judges. Apparently claude code can solve a task with or without a task file just as well, but the existence of this task file makes plans and work more auditable, and not just for bugs, but for intent follow.

Coming back to user intent, I have a post user message hook that writes user messages to a project scoped chat_log.md file, which means all user messages are preserved (user text << agent text, it is efficient), when we start a new task the chat log is checked to see if intent was properly captured. I also use it to recover context across sessions and remember what we did last.

Once every 10-20 tasks I run a retrospective task that inspects all task.md files since last retro and judges how the harness performs and project goes. This can detect things not apparent in task level work, for example when using multiple tasks to implement a more complex feature, or when a subsystem is touched by multiple tasks. I think reflection is the one place where the harness itself and how we use it can be refined.

    claude plugin marketplace add horiacristescu/claude-playbook-plugin

    source at https://github.com/horiacristescu/claude-playbook-plugin/tree/main

Prior to Anthropic I have had bad experiences with Windsurf and Cursor, same shit - I pay the plan, they shrink my usage quota after a short time, couple of months or weeks. I never returned to Windsurf after they abused me, and never used Cursor after I got my Claude sub, I have no idea where I'll end up next. Too bad Anthropic is pushing my $200/mo away.

In a loopy recursive way, it is. Cost gates what we can do and become. Paying back your costs to extend your runway is the working principle behind biology, economy and technology. I am not saying rich people are always right, just that cost is not so irrelevant to everything else. I personally think cost satisfaction explains multiple levels, from biology up.

Related to introspection - it certainly has a cost for doing it, and a cost for not doing it. Going happy go lucky is not necessarily optimal, experience was expensive to gain, not using it at all is a big loss. Being paralyzed by rumination is also not optimal, we have to act in time, we can't delay and if we do, it comes out differently.

I checked my logs and I write 10 words in chat for 1 word in LLM output for final text. So it's clearly not making me type less. I used to type about 10K words per month now I type 50-100K words per month (LLM chat is the difference).

The surplus capacity provided by LLMs got reinvested immediately in scope and depth expansion. I did not get to spend 10x less time writing.

You can own the works, but not the vibes. If everyone owned the vibes we would all be infringing others. In my view abstractions should not be protected by copyright, only expression, currently the abstraction-filtration-comparison standard (AFC) protects abstractions too, non-literal infringement is a thing.

Trying to own the vibes is like trying to own the functionality itself, no matter the distinct implementation details, and this is closer to patents than copyrights. But patents get researched for prior art and have limited duration, copyright is automatic and almost infinite duration.

Anyone wondered if ARC is a measure of intelligence or just a collection of hand picked tasks? was there a proof they encode anything meaningful about intelligence in such short tasks in miniature environments? One shot intelligence?