Communication style

Always explain things simply, in plain language, so I can easily understand. Avoid jargon; when a technical term is unavoidable, define it in a few plain words. Prefer short sentences, concrete everyday analogies, and small examples over dense or abstract explanations. Lead with the plain-English takeaway, then add detail only if it helps.

Git workflow

Don't create branches unless I ask. Commit on whatever branch is currently checked out (the working branch) — even if that's main or the default branch. Only branch first when I explicitly ask for a branch or a PR. This overrides any default "branch first on the default branch" behavior.

Never save screenshots into my project directory. Every screenshot — Playwright MCP (browser_take_screenshot), mobile MCP (mobile_save_screenshot, mobile_take_screenshot), simulator captures, or anything else — must be written to the session scratchpad directory given in your environment prompt (the /private/tmp/claude-*/.../scratchpad path), never to the repo root, .playwright-mcp/, or any folder inside the project. Pass an explicit absolute path under that scratchpad on every call; don't rely on the tool's default output location, because the defaults drop files into the current working directory.
The same goes for any other throwaway file — scratch scripts, intermediate data, debug dumps, generated reports I didn't ask to keep. Only write inside the project when the file is real source I asked for.
If you notice stray screenshots or scratch files already sitting in a project, tell me instead of silently leaving them there.

Always use the Opus model for writing code. Any agent whose job is to write, edit, or refactor code — whether spawned via the Agent tool or as a step in a multi-agent workflow — must use the Opus model. This does not apply to fork agents, which always inherit my model anyway.
Use Opus for code research and exploration too. Any agent that searches, reads, or traces through a codebase must run on the Opus model and return only a concise summary of its findings — always including the concrete file paths, line numbers, and the relevant code snippets/sections, so no re-reading is needed. Fable should never spend tokens grepping, reading files, or exploring code itself — delegate that to an Opus agent and work from the summary it brings back.
Use Fable only for planning and general instructions. Keep Fable for planning, orchestration, and deciding what to do with the summaries agents return. Never use Fable to write code or run codebase searches — delegate both to Opus.

graphify (~/.claude/skills/graphify/SKILL.md) - any input to knowledge graph. Trigger: /graphify When the user types /graphify, use the installed graphify skill or instructions before doing anything else.