AI Agent Workflow Collection

Audit recurring workflows and installed skills to recommend evidence-backed skill updates, additions, or removals, based on repeated procedures rather than recurring topics.

Select and apply the shortest reporting profile, compact, normal, or explicit, that still preserves evidence in a progress update, answer, or handoff. Compresses presentation, never evidence.

Define, implement, redesign, or audit deliberate web interface direction by turning visual intent into inspectable constraints and then verifying the rendered result.

Find the cause of a defect through observable evidence: build a runnable signal, reproduce, minimize, form falsifiable hypotheses, instrument, and prove the fix with regression evidence.

Build or audit a version-neutral prompt contract with explicit outcomes, boundaries, permissions, named tools, required evidence, and stop conditions, then render it as a stable prompt.

Execute long or compaction-prone work with external state and evidence receipts. Use for resumable multi-turn work with several observable criteria; abstain from focused single-pass tasks.

Discover unfamiliar or changing browser workflows through an already-installed Skyvern, writing normalized hashed artifacts. Discovery only, never the final judge of browser behavior.

Land a change in code that has no tests without altering what it silently does - locate the change points, find a seam where behavior can be sensed or substituted, break the dependency with sprout and wrap moves so the class becomes testable, pin current behavior with characterization tests, then make the change inside that safety net. Use when the task involves editing untested code, a module with no coverage that everyone avoids, or a change that must not break callers nobody can enumerate.

Operate and verify native Windows applications through an already-installed QCU, using guarded desktop observation and input that produces hashed evidence.

Refactor existing code for readability and maintainability - naming, function size, nesting depth, duplication, comments, error handling and test clarity - without changing observable behaviour. Use this skill when the code already works and the cost is comprehension: cleanup requests, files described as messy or hard to follow, dead code and stale comments, or when a change lands in a function that is long, deeply nested or duplicated.

Run reproducible research: freeze a pre-registration of question, hypotheses, falsifiers, and stopping rules, run labeled confirmatory and exploratory experiments, then bind every claim to evidence.

Use for cross-file implementation, reproduced defects, bounded primary-source decisions, or authorized real-host workflows. Do not use for obvious one-file edits.

Retrieve current, version-matched authoritative documentation for an external library, framework, SDK, CLI, or cloud API, after establishing the dependency release the repository actually installs.

Audit an already-completed delivery against real evidence. Use when asked to review an existing completion claim, tests, Git state, readiness, or benchmark; never while implementing a change.

Verify known browser behavior or regressions through an already-configured Playwright installation, capturing machine-readable evidence at the public user-visible seam.