workbench

Attacks finished work by trying to disprove what it claims, and reports what survives with the evidence. Use when the user says "adversarial review", "attack this", "try to break this", "poke holes in this", "prove me wrong", "be skeptical", or wants a hostile pass over work that is claimed done — before declaring it finished, shipping it, merging it, or trusting a result.

Charts an effort too big for one agent session as a shared map of decision tickets on the repo's issue tracker, worked one at a time until the way to the destination is clear. Use when the user explicitly invokes /blueprint, says "chart a map" or "work the map", names an existing map issue to continue, or asks to plan something too large and foggy to spec in one sitting. Planning only — it produces decisions, not build slices.

Shape a fuzzy idea into a committed direction before any plan, PRD, or code exists. Use when the user wants to brainstorm, explore approaches to a problem, kick around a half-formed idea, or has a goal but no committed solution yet.

Applies chart-design taste to React data visualization — a chart-type decision tree and adjustable dials (annotation density, complexity, color restraint) to stop charts from being technically-rendered-but-uninformative. Use when building charts or data visualizations with Recharts, Nivo, or similar React charting libraries.

Git commit workflow with enforced conventional commit style. Use when Claude needs to stage and commit changes, craft commit messages, or the user asks to commit, make a commit, or save their work. Ensures consistent commit message format, proper scoping, and atomic commits across the project.

Create and register Claude Code hooks (PreToolUse, PostToolUse) as Python scripts. Use when user wants to create a hook, add a pre-edit check, post-edit formatter, block file edits, or automate responses to tool use.

AI-powered code quality analysis for Python, Markdown, and Mermaid diagrams. Use when the user asks for a "code review" or "quality check"; when Python files need checking for PEP8 violations, unused code, missing type hints, docstring problems, complexity issues, or potential runtime errors; when Markdown documentation needs validation for broken links, heading structure, or formatting; when Mermaid diagram syntax needs validation; or when the user pastes code snippets for analysis.

Expert guidance for working with Dagster and the dg CLI. ALWAYS use before doing any task that requires knowledge specific to Dagster, or that references assets, materialization, components, data tools or data pipelines. Common tasks may include creating a new project, adding new definitions, understanding the current project structure, answering general questions about the codebase (finding asset, schedule, sensor, component or job definitions), debugging issues, or providing deep information about a specific Dagster concept.

Generate a handoff-ready data discovery document for a Snowflake schema or dbt project. Produces a single markdown file with table inventory, coverage summary, Mermaid relationship diagram, and runnable discovery queries with plain English headers. Adaptive depth for engineers or analysts. Use when: onboarding someone to a dataset, documenting available data, creating a data walkthrough, exploring what's in a schema. Triggers: data discovery, discover data, what data is available, document this schema, data walkthrough, schema handoff, onboard to dataset, what tables exist, explore this data.

Expert guidance for working with dbt Core. ALWAYS use before doing any task that requires knowledge specific to dbt, including building or modifying models, writing SQL transformations, configuring tests, running dbt CLI commands, or working with dbt project structure. Common triggers include references to dbt, models, ref(), source(), materializations, seeds, snapshots, dbt build/run/test, YAML schema files, or analytics engineering patterns.

Answers structural questions about a dbt project from its parsed manifest.json rather than from model comments, README prose, or memory. Use when stating or checking how many models, tests, seeds, or marts a project has; when documenting or auditing a dbt repo; when asking which models a seed or source actually feeds, or what a model really depends on; when hunting for seeds and models that nothing references; when confirming what key or grain a model declares; or when a claim about a dbt project's structure needs evidence behind it.

Generate multiple radically different interface designs for a module using parallel sub-agents. Use when user wants to design an API, explore interface options, compare module shapes, or mentions "design it twice".

Verify a test suite would actually catch the bug it claims to prevent, by re-injecting the defect and checking the suite goes red. Use after writing tests for a security boundary, a validator, a detector, or any guard whose failure mode is silent — and before trusting a green suite as evidence.

Use when user says "dev cycle", "development workflow", "full development pipeline", or invokes /dev-cycle to take a GitHub-issues-driven feature from brainstorm through a merged PR. This is the interactive lane — for greenfield work, skill-building, or changes too coupled to delegate.

Production Python coding standards with automatic version detection (3.10-3.13). Use when writing, reviewing, or refactoring Python to ensure adherence to modern type syntax, LBYL exception handling, pathlib operations, ABC-based interfaces, and production-tested patterns. Not Dagster-specific - applies to any Python project.

Build a queue of filed, specced issues to empty by hand, one isolated worker per issue, with an adversarial spec gate before each build and a review of every diff before it lands. Use when several ready issues must be built and merged as a batch, when an unattended executor is unavailable and its backlog still has to move, or when the user says "work the queue", "drain the backlog", "land these tickets", or "build these by hand".

Use when implementation is complete, all tests pass, and you need to decide how to integrate a finished development branch — merge, open a PR, keep it, or discard it. Do not use mid-implementation or while the test suite is red.

GitHub CLI (gh) integration for managing issues, pull requests, branches, commits, and code reviews directly from the terminal. Use when Claude needs to create, list, view, or update GitHub issues; create draft branches and pull requests; make commits and push changes; review pull request diffs and changes; approve or merge PRs; manage GitHub Actions workflows; or work with GitHub repositories without switching to a browser. Requires gh CLI installed and authenticated.

Watch GitLab CI in the background until a pushed commit, a merging MR, or an integration branch head reaches a terminal state, reporting every job's status — roll-up success is never the report. Use after any push to a work GitLab repo (the verify-ci-green rule), after `glab mr merge` returns 405 or flips to auto-merge, or when post-merge CI on dev must be confirmed green. For browsing pipelines, jobs, or logs interactively, use gitlab-cli.

GitLab CLI (glab) integration for managing issues, branches, merge request review, and CI/CD pipelines from the terminal. Use when Claude needs to create, list, view, or update GitLab issues; push branches; review merge request diffs; approve or merge MRs; or inspect and retry pipelines and job logs. Requires the glab CLI installed and authenticated. For CREATING a merge request, use the gitlab-mr-create skill instead; for WATCHING CI to completion after a push or merge, use the gitlab-ci-watch skill instead.

Create GitLab merge requests with `glab` using the `HEAD` conventional-commit subject as the exact title, a Markdown description file with real newlines, and API read-back verification. Use whenever creating a GitLab merge request.

Integration and promotion policy for Clearway GitLab data repos (Dagster, dbt, ingestion). Use when starting work, choosing a branch or merge target, opening a merge request into dev, promoting dev to main, or releasing to production in one of these repos.

Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".

Use when several MRs or PRs are open against the same branch and the user asks which to merge first, whether one blocks another, why merging one breaks another, or in what order to land a queue. Computes a pairwise conflict matrix with git merge-tree and recommends an order by rebase cost. Read-only. Covers GitLab and GitHub.

Use when a user says an MR, PR, merge request, or pull request has review feedback, review comments, changes requested, an approval blocker, or asks to see what needs to be fixed, answered, or replied to after review. Covers reading review threads, triaging findings, landing the fix, and replying to the reviewer in-thread on GitLab and GitHub.

CEO/founder-mode review that rethinks a plan to find the 10-star product. Use when the user asks for a plan review, CEO review, mega review, or wants a plan challenged or stress-tested before implementation.

Break a PRD into independently-grabbable GitHub issues using tracer-bullet vertical slices, with executor-ready issue bodies an autonomous agent can build from directly. Use when user wants to convert a PRD to issues, create implementation tickets, or break down a PRD into work items.

Turn a PRD into a multi-phase implementation plan using tracer-bullet vertical slices, saved as a local Markdown file in docs/plans/. Use when user wants to break down a PRD, create an implementation plan, plan phases from a PRD, or mentions "tracer bullets".

Generate or update the `.claude/docs/project.md` file that gives Claude project-specific context. Use this skill when the user asks to create, update, regenerate, or refresh project context, or says things like "update project.md", "generate project context", "this repo needs a project.md", or "Claude doesn't know about this project". Also trigger when onboarding Claude to a new repository for the first time.

Applies deliberate design taste to React UI generation — adjustable dials (variance, motion, density) and explicit anti-genericness rules to stop AI-generated components from defaulting to the generic shadcn/Tailwind look. Use when building or editing React components, pages, or layouts (.tsx/.jsx), especially new UI generation in Tailwind/shadcn/Framer-Motion projects.

Creates, classifies, and maintains a repository's human-facing documentation as one Diátaxis-shaped set: the root README landing page and docs/ split into tutorials, how-to guides, reference, and explanation, with a provenance footer and a drift, link, and mode checker. Use when someone asks to write, generate, update, improve, or refresh a README, says "this repo needs a README", wants a runbook, how-to guide, operations guide, tutorial, architecture doc, module map, data-flow write-up, or a "where does X live" doc, asks which kind of doc something is or whether a page is a how-to or reference, wants docs restructured or migrated under docs/, or wants a staleness or link check of existing docs against the code. Not for the Claude-facing project.md (use project-context) or docs/plans/ (use prd-to-plan).

Use when user wants to plan a refactor, create a refactoring RFC, break a refactor into safe incremental steps, or find architectural improvement opportunities (deepening shallow modules, consolidating tightly-coupled code, making a codebase more testable or AI-navigable). Produces a detailed refactor plan with tiny commits, filed as a GitHub issue.

Security code review for vulnerabilities with confidence-based reporting. Use when the user asks for "security review", "find vulnerabilities", "check for security issues", "audit security", "OWASP review", or to review code for injection, XSS, authentication, authorization, or cryptography issues.

Set up pre-commit hooks for the current repo. Use when user wants to add pre-commit hooks, configure commit-time linting, formatting, type checking, or testing. Triggers on "pre-commit", "git hooks", "linting hooks", or /setup-pre-commit.

Protect work when a git working tree or worktree may be shared with a live autonomous agent or another session. Use before resetting, force-checkouting, or cleaning any tree an agent might be using, when a working tree changes unexpectedly mid-task, or when taking over a directory another process was working in.

Compile-check committed warehouse SQL (Snowflake, BigQuery, Redshift) against the live schema before deploying it, catching column drift and views that will not build. Use when about to run a .sql file at a warehouse, after any ALTER TABLE, before re-running a views file, or when a deploy half-applied and left objects inconsistent. Skip for local-only or fixture SQL.

Use before acting on any recorded artifact — an issue, a review finding, a "do not merge" comment, a TODO or blocker doc, a plan prerequisite, a branch someone said still needs reviving. Re-verifies each against current reality and classifies it with evidence. Read-only.

Test-driven development with red-green-refactor loop. Use when user wants to build features or fix bugs using TDD, mentions "red-green-refactor", wants integration tests, or asks for test-first development.

Turn a YouTube lecture/talk or a raw transcript (VTT, SRT, or plain text) into a readable Obsidian-markdown study note — imposed structure, reconstructed LaTeX with plain-word glosses, flagged missing visuals, and per-section reading prompts. Trigger when the user gives a YouTube URL/video id to ingest, or uploads/pastes a transcript and wants it made readable ("ingest this talk", "clean up this transcript", "turn this lecture into notes"). Handles lectures, talks, webinars, interviews, and panels. Do NOT trigger to summarize (this skill never summarizes), to quiz/drill, or to typeset non-transcript prose.

Use when user reports a bug, wants to file an issue, mentions "triage", or wants to investigate and plan a fix for a problem. Finds the root cause and files a GitHub issue with a TDD-based fix plan.

Diagnose and resolve a failed, quarantined, or question-parked autonomous agent run, reusing its preserved work instead of rebuilding. Use when an unattended agent slice quarantines or asks a question, a nightly agent run exits nonzero, a background agent's work was rejected by a gate or reviewer, or the user asks why an automated run failed or what to do with parked work.

Use when starting any conversation or task in this project — establishes precedence between instructions and skills, requires invoking any skill that might apply, and sets the order skills run in before any response or action.

Run Charles's vault (The Vault) /vault-audit structural audit across frontmatter, wikilinks, indexes, stale notes, duplicates, and templates. Trigger when Charles invokes /vault-audit, mentions /vault-audit, or asks for this vault workflow by name.

Run Charles's vault (The Vault) /budget spend and subscription-value meter from local Claude transcripts. Trigger when Charles invokes /budget, mentions /budget, or asks for this vault workflow by name.

Run Charles's vault (The Vault) /clickup-task-sync workflow to sync vault action items into ClickUp without duplicating tasks. Trigger when Charles invokes /clickup-task-sync, mentions /clickup-task-sync, or asks to sync action items / a 1:1 recap into ClickUp.

Run Charles's vault (The Vault) /cold-read gate — an adversarial read of a dispatched issue's SPEC (not its code) before it is promoted to the afk executor. Trigger when Charles invokes /cold-read, mentions /cold-read, asks whether an issue is buildable cold, or is about to promote a `proposed` issue.

Run Charles's vault (The Vault) /connect autonomous graph connection pass with preview-gated wikilink edits. Trigger when Charles invokes /connect, mentions /connect, or asks for this vault workflow by name.

Run Charles's vault (The Vault) /context-then-delegate workflow to resolve real-world ambiguity (email/SharePoint/Slack) before writing a coding-agent prompt. Trigger when Charles invokes /context-then-delegate, mentions /context-then-delegate, or is about to write a delegated prompt for a task with unresolved domain ambiguity.

Run Charles's vault (The Vault) /debrief retrospective over recent afk builds. Trigger when Charles invokes /debrief, mentions /debrief, asks how much of the autonomous work needed his hands, asks whether /cold-read is actually earning its cost, or asks which stage keeps producing rework.

Run Charles's vault (The Vault) /dispatch workflow to turn a shaped idea into an afk-managed issue linked back into the vault. Trigger when Charles invokes /dispatch, mentions /dispatch, or asks for this vault workflow by name.