13 lines
1.0 KiB
Markdown
13 lines
1.0 KiB
Markdown
# Root Skill Reference
|
|
|
|
## Why the State Machine?
|
|
LLMs are prone to "momentum", where they output a plan and immediately start executing it in the same breath, often writing the wrong files or breaking things before the user can review.
|
|
The `PLAN` vs `IMPLEMENTATION` state machine forcefully stops the AI between thought and action.
|
|
|
|
## Why the Active Skill Echo?
|
|
Context windows shift. If you read a skill 5 turns ago, you will forget the strict prohibitions (`🚨 Prohibitions`).
|
|
By forcing you to output `### Active Skills:` and physically writing out the 2-3 most critical rules, the LLM attention mechanism (Self-Attention) locks onto those rules for the duration of the generation step. It is a hack to prevent "lost in the middle" syndrome.
|
|
|
|
## Migration History
|
|
These root instructions previously lived in the top-level `.ai/skills/root-skill/SKILL.md` file but were migrated into the Matt Pocock Agent Skills standard (`.ai/skills/root-skill/SKILL.md`) to ensure a 100% unified architecture. Now, even the root file is just a skill.
|