oberskills

Subagent dispatch guidance for the Agent tool — whether to delegate or work inline, the four-part delegation contract, model and effort selection, parallel fan-out sizing, fork vs fresh subagent, and dispatching independent verifier agents. Use before any Agent tool call: spawning subagents, fanning out work across files or research angles, delegating exploration or review, briefing a verifier, or deciding whether delegation is worth it at all. Not for: wording a prompt that isn't an agent dispatch (use oberskills:prompt) or authoring reusable skill and agent definition files — structure, frontmatter, and evals (use oberskills:skill-craft); their prompt bodies (use oberskills:prompt).

Controls a live Chrome browser through a persistent puppeteer-core MCP connection — click, type, hover, drag, fill forms; navigate pages; take screenshots and export PDFs; read the DOM and accessibility tree; run Lighthouse audits and Core Web Vitals traces; intercept, record, stub, or block network requests; export HAR traffic; emulate devices and network conditions; manage browser storage; upload and download files. Use when interacting with or automating a running browser, clicking elements, taking screenshots, running Lighthouse, capturing network traffic as HAR, checking the accessibility tree, intercepting or stubbing requests, or emulating a device or viewport. Not for: web search or fetching public URLs without a browser session (use web-research), reading a static HTML file already on disk, fixing TypeScript or JavaScript build or compile errors, debugging source code without a running page, or tasks that don't require controlling a live browser.

Design and review prompts Claude-first — system prompts, reusable agent definitions, long or novel dispatch briefs, pipeline stages, and prompts users will run elsewhere. Two modes: design and adversarial review. Covers Claude model behavior and migration, de-prompting, few-shot design, output-schema ordering, context engineering, prompt security, optimization, non-Claude/Codex target review via references/porting.md in this skill directory, and a verbatim behavior-snippet library. Use when writing, improving, debugging, reviewing, or migrating any prompt. Not for: creating, reviewing, or testing skill files (use skill-craft); routine subagent delegation prompts (use agent).

Captures and summarizes screenshots from the local desktop or a named window using the bundled capture script. Use when the user asks for a screenshot, screen capture, active-window capture, visual inspection of the current screen, or named-window screenshot. Not for browser-page screenshots when browser MCP tools are already controlling a page.

Create, evaluate, and review Claude Code skills and reusable agent definition files (structure, frontmatter, evals). Covers the skill-vs-subagent-vs-hook decision, SKILL.md authoring and frontmatter, trigger-description optimization, baseline-first evals with pressure testing via the skill-eval MCP tools, validation, and packaging. Use when creating a new skill, improving or benchmarking an existing one, reviewing a skill directory, or writing evals for a skill. Not for prompt wording design or review — including the prompt body inside an agent definition (use oberskills:prompt), live subagent dispatch (use oberskills:agent), or prose editing (use oberskills:write).

Runs parallel multi-angle web research at four depth modes (scan, brief, breadth, deep): a planner grounds the query in local context and plans search dimensions, parallel search agents pull verbatim extracts with source URLs, one dimension always hunts counter-evidence, and synthesis is grounded back in the local project. Use when researching a topic, library, or tool on the web, comparing options or products, surveying a landscape of approaches, or checking current, version-sensitive, or time-sensitive information. Not for: driving or automating a live browser session (use oberskills:browser), academic literature search across arXiv or Semantic Scholar, or questions about the current repo, project, or files already on disk — read those directly.

Edits and reviews prose to remove the AI tells a model cannot see in its own writing — measured detection signals, not intuition: structural rhythm first (sentence-length variance, information density, discourse order), the surface kill-list second (banned phrases, em-dash overuse, hollow openers, hedging), with an optional phased interactive review for longer pieces. Use whenever prose for human readers is being written or edited — blog posts, announcements, docs, emails, READMEs — when text needs humanizing or sounds robotic, and when reviewing someone's writing. Not for: wording prompts or system-prompt text (use oberskills:prompt), converting markdown for platforms like Slack or Notion (use penman), or writing code, comments, docstrings, or API reference docs.

Decomposes user intent through structured brainstorming before acting on ambiguous or underspecified requests — classifies the ambiguity type, generates competing interpretations, and selects the clarifying question that removes the most uncertainty. Use when requirements are unclear, a request could have multiple valid interpretations, critical details or constraints are missing, or the request contains contradictions. Most requests need no clarification — invoke only when a concrete ambiguity would change the approach. Not for: mechanical or well-specified tasks (renames, typo fixes, version bumps, running tests or commands, any edit with an explicit file and target), details you can determine yourself from the codebase, or full research or planning workflows that own their own scoping.

Rewrite any reply in pirate speak. Use whenever the user asks to talk like a pirate or mentions pirates.