Agent skills

Let your agents inspect, search, and evaluate their own history.

thirdeye ships four Claude-Code-format skills so the agents you trace can also use their history. Installed once into a repo, they're discoverable by Claude Code, Cursor, and any other client that reads the .agents/skills/ or .claude/skills/ conventions.

The bundled skills

use-thirdeye

Basic CLI fluency. Teaches an agent to:

  • Enable tracing for a platform (thirdeye add).
  • Search and retrieve sessions (thirdeye list, thirdeye events, thirdeye event, thirdeye tail, thirdeye search).
  • Debug failed tool calls by filtering events and reading payloads.
  • Read per-turn token usage from thirdeye usage.
  • Tag and curate events (thirdeye tag, thirdeye tags).

Reach for this skill when you want your agent to introspect a past session — yours or its own — to debug, summarize, or follow up on prior work.

thirdeye-evals

The full eval workflow. Teaches an agent to:

  • List and show shipped rubrics; create custom ones with thirdeye eval def create.
  • Dispatch evaluators with thirdeye eval run (foreground or --background).
  • Read results with thirdeye eval show / thirdeye eval list.
  • Interpret per-turn findings anchored to event seq.
  • Apply common patterns (token efficiency, tool quality, error recovery, task adherence, redundancy).

Pair it with use-thirdeye — the eval skill assumes you can already navigate sessions.

thirdeye-review

A workflow for auditing other agents' traces to find inefficiencies and propose fixes. Teaches an agent to:

  • Scope a review cohort by agent type, repo, and time window.
  • Cluster sessions by the task they were performing.
  • Audit tool choices and execution patterns against known waste signatures (re-reads, bash-grep, retry loops, unbounded payloads).
  • Quantify the token cost or error rate behind each pattern.
  • Tie each finding back to its invocation surface — prompt, skill description, conventions file, plan task, or tool allowlist — and propose a concrete edit.

Use this when asked to reduce agent costs, improve results, or suggest new skills based on what you observe in past traces. Output is a structured report whose recommendations are each tied to a specific file or flag.

thirdeye-filter

A directive used by the browser UI's Ask panel to translate natural-language queries ("long-lasting claude runs this week") into the structured filter JSON the UI applies to your sessions or search view. Installed automatically with the other skills; not normally invoked directly by you or your agents.

Installing into a repo

thirdeye skill list                                  # show bundled skill names
thirdeye skill install                               # all skills → .agents/skills/
thirdeye skill install .claude/skills                # custom parent dir
thirdeye skill install --only thirdeye-review        # just one
thirdeye skill install --force                       # replace existing entries

By default thirdeye skill install writes into .agents/skills/ for cross-client discoverability. Pass a path argument to target a different convention (e.g. .claude/skills if you only use Claude Code).

Symlinks, not copies

Skills install as symlinks pointing at the bundled skill content inside the installed thirdeye package. So when you upgrade thirdeye (brew upgrade thirdeye or pipx upgrade thrdi), the skill content refreshes automatically in every repo that has them installed — no per-repo reinstall required.

If you want a frozen copy that won't move with thirdeye upgrades, install once and then replace the symlink with cp -RL.

Using a skill

Claude Code

After install, Claude Code picks up skills under .agents/skills/ and .claude/skills/ automatically. Mention the skill by name in your prompt — for example, "use the use-thirdeye skill to find sessions that touched auth.py last week" — and Claude will load the skill instructions.

Cursor / Codex / Gemini

Any client that reads the .agents/skills/ convention will discover the same files. For clients that don't, point them at the SKILL.md directly — the file is plain markdown and the rest of the skill (the references/ directory) is just supplementary content the agent reads when needed.

Tips

  • Install them all together unless you specifically want a subset — agents tend to chain "find a problem" → "grade the problem" → "review the cohort" naturally once the surfaces are on hand.
  • Re-run thirdeye skill install if you've moved or renamed your skills directory; it's idempotent and --force will replace stale entries.
  • Audit what the skill does. Skills are markdown — read them. They're at .agents/skills/use-thirdeye/SKILL.md (or wherever you installed them) and link out to the reference files for each workflow.