Custom Investigation Selector
Lets an agent review a chosen slice of agent conversations via a SQL selector instead of a random sample.
- Repository
- google/adk-recipes
- Stars
- 10,568
- Licence
- Apache-2.0
- Files
- 4 files
What it does
This skill teaches an AI agent to run a targeted review of another agent's recorded conversations. Instead of a random sample, the agent writes a SQL query (a 'selector') over telemetry data to pick conversations matching criteria such as errors, a specific tool, slow latency, a topic, looping, or abandonment. It previews the matches with the user, then starts the review, and can also narrow the reviewer's focus.
Great for
- Developers running the ambient quality agent who want to examine specific failing or unusual conversations
- Teams debugging a particular tool, error pattern, or latency problem in a deployed agent
- Engineers re-reviewing past conversations with a different focus
Skip it if
- Deployments that score sessions rather than review them: the tools refuse outright in that mode, so the skill cannot run there
- Users without access to the observed agent's telemetry tables, since the selector can only read that one table
- Anyone expecting the preview examples to be the exact conversations reviewed, because preview and run take separate random samples
Try asking
- “Investigate only conversations that hit an error in the last 24 hours”
- “Review conversations where the search tool was called more than six times”
- “Find sessions that took longer than two minutes and focus the reviewer on latency”
- “Re-review conversations that mention billing with a focus on abandonment”
Install in 3 steps
- 1
Download the skill
The ZIP contains the 4 original files, the Apache-2.0 licence and an attribution note.
custom-investigation.zip - 2
Add it to your agent
Claude Code (all projects)
mkdir -p ~/.claude/skills && unzip custom-investigation.zip -d ~/.claude/skills/Claude Code / Agent SDK (this project only)
mkdir -p .claude/skills && unzip custom-investigation.zip -d .claude/skills/Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.
- 3
Ask for the task
No special command needed — the agent reads the skill's description and loads it when your request matches. Try one of the prompts above.
Get better results
- Let the agent preview first and open the example links yourself before approving a run, since the examples show whether the selector picked the right conversations.
- Keep the selector narrow but use HAVING for group-level conditions like repeated tool calls or total duration.
- If a preview is rejected three times, review the explanation yourself rather than asking the agent to keep rewriting the SQL.
What's inside
- references/big_query.md
- references/cloud_logging.md
- references/cloud_ops.md
- SKILL.md
Redistributed unmodified from google/adk-recipes under the Apache-2.0 licence. The assessment was drafted with AI from the skill's files on 2026-10-29. Skills can instruct your agent to run code — read SKILL.md and any scripts before installing.