Haize's collection of agent skills, following the Agent Skills format.
Skills are packaged instructions, workflows, and reference materials that help coding agents handle specialized tasks more reliably. This repository is where we keep project-specific skills that encode investigation methodology, reusable patterns, and working artifacts.
Investigate agent behavior from traces, logs, trajectory stores, and observability backends.
Use when:
- analyzing agent failures, stalls, regressions, or slowness
- auditing reasoning quality or behavioral consistency
- comparing behavior across models, runtimes, or agent versions
- understanding multi-agent coordination or interaction dynamics
- separating tool failure from tool misuse
- turning raw traces into a structured investigation report
- setting up or improving tracing for future investigations
What it includes:
- a schema-first, hypothesis-first investigation workflow
- runbooks for operational, reasoning, social, and safety behaviors
- self-linting guidance to catch shallow or narrative-driven investigations
- HTML visualization templates for report outputs
- tracing references and investigation documentation
Install the skill with:
npx skills add https://github.com/haizelabs/skills/tree/main/agent-trace-investigatorSkills in this repository are meant to be available to the agent when working in this codebase.
Typical prompts look like:
Investigate these traces and tell me what the agent thinks before it submits its final work.
Compare model behavior across these two agent versions.
Add or update a skill when there is project knowledge that should become reusable agent behavior.
Good additions include:
- a repeatable workflow the agent should follow
- a reference guide that improves output quality
- a reusable investigation or debugging pattern
- concrete examples, templates, or guardrails
When updating an existing skill, prefer improving the current docs and references over duplicating concepts across multiple places.