Use when bootstrapping or overhauling an agent system by dispatching investigative tracer jobs and harvesting friction notes for a consolidated solid-loop pass.
Installs into .claude/skills of the current project.
Are you the author of Boot Loop?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/wayne930242-boot-loop)
---
name: boot-loop
description: Use when bootstrapping or overhauling an agent system by dispatching investigative tracer jobs and harvesting friction notes for a consolidated solid-loop pass.
---
# Boot Loop
Bootstrap or overhaul a project's agent system from empirical evidence gathered by parallel investigative workers.
`boot-loop` probes the codebase through focused investigation tasks, collects the friction notes they produce, and runs one `solid-loop` pass on them.
## Workflow
### 1. Scope Investigative Tracer Jobs
Select 2 to 4 distinct, high-impact investigation targets representing primary areas of the repository:
- **Build & Test Reality**: Run test suites, typecheckers, or linters to identify local execution quirks and missing setup steps.
- **Architectural Seams**: Trace a core user flow or data pipeline to identify primary entry points and boundaries.
- **Conventions & Idioms**: Inspect typical module structure, configuration conventions, and dependency patterns.
Each job must have a concrete, falsifiable objective.
### 2. Dispatch Independent Workers (`investigating`)
Dispatch isolated subagents or background workers for each target:
- Each worker follows **`investigating`**.
- Each worker records and classifies friction notes in the `aaaav-do` format, then returns the classified notes instead of invoking `solid-loop`.
Wait for workers to complete their designated investigation runs.
### 3. Merge Friction Notes
Collect the classified notes from all completed workers and merge notes whose `Found` lines state the same fact.
### 4. Run One `solid-loop` Pass
Apply the merged notes through a single `solid-loop` pass.
### 5. Verify Agent System Integrity
1. Run the workspace validator (`uv run python hooks/validate_all.py .`) to ensure:
- Frontmatter and trigger phrasing are valid.
- All skill bodies remain under 300 lines with valid local links.
- Expected behaviors are stated directly without defensive phrasing.
2. Emit a concise completion summary listing:
- Workers dispatched and jobs executed.
- Friction notes applied and their actions.
- Updated or created agent system components.