dots

Consult a stronger reviewer who sees your full conversation transcript.

Design a consequential new or changed code boundary before implementation when current requirements force decisions about ownership, lifecycle, contracts, atomicity, or state. Use for 'architect this,' new APIs or module boundaries, and changes whose types, ownership, or state model must be settled first.

Audit a codebase or subsystem for structural refactors, misplaced ownership, duplicated policy, weak boundaries, and code to consolidate or delete. Use for broad architecture and cleanup reviews, not explanation-only requests or reviews limited to defects introduced by a selected change.

Keeps a GitHub pull request moving by watching its latest checks and review feedback, fixing problems within its original goal, and replying to handled feedback. Use when the user asks to monitor, watch, or babysit a pull request; not for a one-time status report.

Use only when the user selects `$clarify` to resolve underspecified requirements before implementation; not for shaping a fuzzy idea, planning the implementation, or asking questions the repository can answer.

Review completed code changes before merging for correctness, readability, simplicity, and maintainability. Returns findings by default; repairs retained in-scope findings when invoked by an authorized implementation workflow, otherwise only when the user explicitly asks.

Use when creating, redesigning, implementing, or polishing visible web or app UI, including screens, flows, components, responsive layouts, and design-system work. Makes product changes; for an independent read-only critique or ship-readiness assessment, use design-review.

Review existing product UI, screenshots, prototypes, or flows for visual quality, accessibility, design-system conformance, motion, prototype-to-implementation fidelity, and ship readiness. Returns actionable findings without changing product source; use design to create, implement, redesign, or polish UI.

Use when writing, revising, or reviewing durable repository documentation such as READMEs, how-tos, runbooks, design docs, ADRs, troubleshooting guides, changelogs, release notes, PRDs, and agent-readable knowledge. Not for project instructions, code comments, chat explanations, publishing, or broad implementation.

Use only when the user writes `$explain`. Give a plain-language or ELI5 explanation of any subject, or the previous answer when none is named, building intuition through examples and diagrams for code-related subjects; not for rigorous code investigation, critique, review, or implementation.

Creates a continuation brief for the current task, branch, pull request, plan, or conversation. Use when the user asks to hand work to another task, agent, or phase; use recall to reconstruct earlier work and pr to publish changes.

Explains how code, a subsystem, change, commit, branch, or pull request works, including runtime flow, ownership, placement, and layering. Use why for historical rationale, explain for a quick ELI5 answer, and architecture-review for a broad structural audit.

Creates and edits self-contained browser-openable HTML pages, linked page sets, fragments, static product mocks, templates, and HTML assets for template-driven skills. Not for production UI implementation, interactive data visualizations, underlying research or planning, durable documentation, or non-HTML skill authoring.

Route work to the smallest applicable Dots skill, shared reference, or Feature Development workflow only when the user explicitly invokes Dots, $index, or Dots Feature Development. Ordinary implementation and bug-fix requests do not trigger this router. Focused requests use their owning skill.

Create, rewrite, or tighten a standalone prompt for another assistant from a task description, existing prompt, or supplied context. Use for prompt-writing requests, including coding-agent and image prompts; not for performing the task described in the prompt or explaining prompting techniques.

Coordinate agents on substantial tasks. Use when the user requests subagents or parallel work would improve speed or breadth; not for trivial work or workflows that already define their own team.

Commits and pushes finished changes, then opens or updates a non-draft GitHub pull request ready for review. Use when the user asks to publish completed work as a PR. Not for merging, addressing review feedback, or monitoring checks and reviews.

Builds an isolated throwaway prototype to resolve a specific behavioral, technical, interaction, timing, or visual choice through observation; not for production implementation or settled static HTML work.

Use only when the user selects `$recall` to reconstruct recent working context, catch up, or find where work stopped. Reads the smallest relevant Codex or Claude task history, optional computer history, shared records, and live state; not for workflow audits or durable self-improvement.

Use only when the user selects `$scout` to be interviewed, challenged, or helped settle a fuzzy plan, decision, or idea before planning or implementation; not for minimal requirements clarification.

Use only when the user selects `$self-improve` to learn from agent work, diagnose workflow or harness friction, evaluate how they work with agents, or find durable improvements across Codex or Claude sessions. Not for static skill review, product fixes, or reconstructing context without improvement.

Creates, updates, or statically reviews agent skill source against Dots standards. Use for authoring, revising, or diagnosing a skill; not for behavioral evaluation or plugin packaging.

Transcribes local media and YouTube videos into Markdown, text, JSON, SRT, or VTT, then creates task-shaped YouTube summaries when requested. Uses captions first and free local speech recognition otherwise. Use for recorded-media transcripts, subtitles, or YouTube summaries; not live capture or paid/cloud transcription.

Use for 'why does X work this way', design rationale, regressions, postmortems, or evidence-backed thresholds. Searches the smallest useful historical record and separates direct evidence, inference, and unknowns. Use how for runtime behavior.