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?
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 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.
the obsession on token discounts recently is pretty funny. If you extrapolate far enough you end up back to where we started, programming languages.
[dead]
So—let me get this straight—this is just a suggestion to format skills' dot-md files with inline `makefile` pseudocode codeblocks for the sake of one-shot alignment?
Have you tried it in Polish? [0]
0. https://arxiv.org/html/2501.02266v1