pulse-data

Investigate a stable counts validation failure for a state supervision or incarceration period table by identifying which fields changed and why. Use when a stable counts validation is failing or has recently failed for a state entity (e.g. "why is the NC stable counts failing", "investigate stable counts for US_ID supervision periods", "dig into the stable counts failure for supervision periods").

Create a GitHub PR for the current branch with auto-generated description. Use when the user asks to create a pull request, open a PR, or put up changes for review.

Investigate a GitHub bug ticket by fetching PII from the go/github-pii doc and suggesting BQ investigation queries.

Post a validation comment to a GitHub PR or issue using the standard format. Use when the user asks to post a comment to a PR or issue, post sandbox or ingest validation results, share validation findings on a PR or ticket, or anything similar that involves posting validation results as a GitHub comment.

Add a new field to opportunity configs in the workflows admin panel. Use when asked to add a new config field, property, or attribute to the workflows opportunity configuration system.

Use this skill when a user wants to structure BigQuery views into the standard Pollen view pattern. All Pollen analysis is in the research/notebooks/policy folder -- do not apply this skill outside that folder. Triggers include: "structure my views", "create view files", "pull views from my notebook", "set up the view folder structure", "make my SQL into view builders", or any request to convert notebook SQL or .sql files into the SimpleBigQueryViewBuilder pattern used by policy analysis work.

File issues and tasks for the pulse-data repository. Use when the user asks to file, create, or track issues, tasks, tickets, or TODOs. Defaults to filing in Linear (which two-way-syncs to GitHub) and lets you set Linear-specific fields like project and milestone; falls back to filing directly in GitHub when Linear is unavailable. Handles templates, labels, team/project/milestone assignment, parent linking, and in-code TODO references.

Create a complete workflow initiative with an epic, core workflow sub-issues, and launch preparation tasks. Use when the user asks to create a workflow epic, file a workflow task, or start a new workflow initiative.

Proactively update SKILL.md files when they become outdated, incorrect, or need improvements. Use when you notice a skill doesn't match current workflow, when user corrects your skill usage, or when discovering better practices that should be documented.

Find illustrative anecdotes from supervision case notes for a given topic. Pulls a filtered batch of raw case notes from BigQuery (using existing document collection configs as the source query), groups by person, and reads through them to surface salient examples. Optionally scopes the candidate set to "Recidiviz-implicated" people — those who interacted with our tools (Workflows opportunity engagement, the Tasks contact-compliance tool, the PSI sentencing assistant pre-sentence, or the JII tablet app) — for impact-attributed storytelling. Use when the user asks for case note anecdotes, real-world examples from case notes, narrative stories from supervision data, qualitative evidence for an impact deck, or wants to back up an analysis with qualitative evidence.

Find people implicated by Recidiviz tools, in two modes. Cohort mode — given a description of a tool-engagement pattern (e.g. "people viewed in Workflows for early discharge AND granted", "people registered for the JII tablet app", "people whose officer was flagged as an outlier during their supervision"), produce a SQL CTE that emits the matching person set. Individual mode — given a `(state_code, person_id)`, summarize which Recidiviz tools implicated them and how (surfaced/viewed/granted, login counts, PSI case view dates, outlier officer flags).

Diagnose a failed Airflow task running on Cloud Composer by pulling task logs from GCS, identifying the first real error in the traceback, and inspecting the task's source and recent git history. Use when the user wants to investigate an Airflow task failure (e.g. "why did X task fail", "investigate the failing generate_airflow_dag_run_history task", "dig into the airflow failure in prod"). Can be invoked directly or as a specialist from the investigate-pd-incident skill.

Investigate a PagerDuty incident by fetching it via the PagerDuty MCP server, classifying the alert type, and dispatching to the appropriate specialist skill. Use when the user provides a PagerDuty incident URL or ID (e.g. "investigate this PD alert", "what's going on with incident Q09D28GV4W96FQ", "diagnose https://recidiviz.pagerduty.com/incidents/...").

Go through a PR's review comments one by one, propose solutions, and commit fixes. Use when the user wants to address PR feedback, review comments, or resolve PR review requests.

Commit uncommitted changes with clear commit messages. Use when the user asks to commit changes, save work, or create a commit.

Upgrade Cloud Composer (Apache Airflow) to a new version. Use when the user asks to upgrade, update, or bump Airflow, Cloud Composer, or Airflow dependencies to a new version. Handles updating terraform config and Python dependencies.

Proactively update CLAUDE.md documentation when it becomes outdated or incorrect. Use when you notice documented commands, paths, or patterns don't match the actual codebase, when making code changes that invalidate documentation, or when user corrects stale information.