“ Dependency Resolution: The harness resolves the DAG (Directed Acyclic Graph) automatically. No more relying on an LLM to "guess" the next logical step.Uses the Target: Dependency + Recipe model to ensure Agents follow a strict execution order without skipping steps.”
How does it do that? Does it just generate a Makefile? If so, why not just put the actual Makefile as a resource in the skill package and provide execution commands? That way the Makefile doesn’t need to be read at all.
If not, and you rely on an LLM interpreting the execution order, wouldn’t that statement just be false?
What I did here was to rewrite SKILL.md in Makefile style, using a DAG structure and omitting the text describing the process. So this should be considered a pseudo-Makefile; writing a SKILL using the Makefile approach is a very natural method.
My next step is to design the recipe to be hot-loadable. The goal is to achieve self-evolving, optimizing the recipe independently without changing the DAG. This ability to perform local optimization is something Markdown lacks, but Makefiles can.
This is interesting. Do you have a robust skill built with this I could checkout? I have been working on a planning skill that has sub-agent that do stuff like research with Tavily and Exa an it uses Claude CLI and Codex CLI to write separate plans and compare and uses a plan template with a micro task layout with multiple phases, test, etc.
it's crazy, I have seen so many projects popping up just focusing on reducing token usage. At least caveman speak is funny!
Have to say that since we switched to our own model in a rented GPU, we stopped worrying about tokens and just use the hell out of our AI as much as we want :)
The core idea of this project is to use Makefiles to style SKILL documentation, leveraging Makefiles' built-in DAG functionality and a defined syntax. The advantages are as follows:
1. It reduces the token consumption of the original MD format;
2. SKILLs are easier to read and more suitable for AI use because the inherent DAG is a Plane Mode;
3. Makefiles are ideal for auditing (git tracing, call statistics), providing a solid fundation for future self-evolving enginering.
I don’t get it.
“ Dependency Resolution: The harness resolves the DAG (Directed Acyclic Graph) automatically. No more relying on an LLM to "guess" the next logical step.Uses the Target: Dependency + Recipe model to ensure Agents follow a strict execution order without skipping steps.”
How does it do that? Does it just generate a Makefile? If so, why not just put the actual Makefile as a resource in the skill package and provide execution commands? That way the Makefile doesn’t need to be read at all.
If not, and you rely on an LLM interpreting the execution order, wouldn’t that statement just be false?
What I did here was to rewrite SKILL.md in Makefile style, using a DAG structure and omitting the text describing the process. So this should be considered a pseudo-Makefile; writing a SKILL using the Makefile approach is a very natural method.
You're just repeating the readme, not answering the question.
My next step is to design the recipe to be hot-loadable. The goal is to achieve self-evolving, optimizing the recipe independently without changing the DAG. This ability to perform local optimization is something Markdown lacks, but Makefiles can.
This is interesting. Do you have a robust skill built with this I could checkout? I have been working on a planning skill that has sub-agent that do stuff like research with Tavily and Exa an it uses Claude CLI and Codex CLI to write separate plans and compare and uses a plan template with a micro task layout with multiple phases, test, etc.
the obsession on token discounts recently is pretty funny. If you extrapolate far enough you end up back to where we started, programming languages.
Yeah, I wonder what kind of work people do that they need more than 500k or 1M context window.
Even when it's a big project, breaking it down doesn't change the output quality.
it's crazy, I have seen so many projects popping up just focusing on reducing token usage. At least caveman speak is funny!
Have to say that since we switched to our own model in a rented GPU, we stopped worrying about tokens and just use the hell out of our AI as much as we want :)
remember TOON? it killed JSON
/s just in case
The thing which is 98% JSON and absolutely didn't kill JSON?
The core idea of this project is to use Makefiles to style SKILL documentation, leveraging Makefiles' built-in DAG functionality and a defined syntax. The advantages are as follows:
1. It reduces the token consumption of the original MD format;
2. SKILLs are easier to read and more suitable for AI use because the inherent DAG is a Plane Mode;
3. Makefiles are ideal for auditing (git tracing, call statistics), providing a solid fundation for future self-evolving enginering.