agent-skills

Create receipt-backed persona dream packets from memory residue. Use when a persona should dream, reflect, or turn recent memories into persona insight; when create-movie/dream.py feels too heavy for the goal; when the desired output is a prompt, frame prompts, contact sheet, reflection, and memory write receipt rather than a full movie; or when a downstream movie workflow needs a dream_packet.json input.

Use when the user asks to query project memory, ask an oracle, use supported browser-backed reviewers, run Tau roundtable/single-handler workflows, run persona/deep-review workflows, generate image prompts, check OS/project health through composed skills, or run an ask DAG. This skill is the executable /ask runtime; do not replace it with an informal subagent, plain web search, or hand-written review.

Build structured evidence cases using Claims-Arguments-Evidence (CAE) trees. AGENT-DRIVEN composable orchestrator: the agent decomposes the question, calls existing skills for data collection and verification, then DECIDES the verdict. Python runner.py is a thin data collector and persistence layer.

Monitor Herdr-visible Codex/Claude agent panes for stalled, blocked, unknown, or confused state, then ask each stalled agent whether it needs human intervention or can self-unblock with brave-search or webgpt.

Internal Tau-owned LLM proxy on localhost:4001. Surfaces: chat/batch completions, scillm exec, OpenCode serve (coding delegate), OpenCode transport (DAG/SSE), standing Codex agents. Chutes, Gemini, Claude/Codex OAuth, OpenCode Go, Ollama. Auto-routes by model name. ZIP/PDF, JSON repair, batch pools. Project agents must not call this skill directly unless the human explicitly asks to operate SciLLM or the work is Tau/SciLLM maintenance.

Operate and verify the local T'au project at ${HOME}/workspace/experiments/tau. Use for Tau loop, harness, watchdog cron, GitHub issue orchestration, TUI, Memory-first chat, and E2E proof/status tasks. This skill is a light wrapper around the Tau repo and must report mocked/live proof boundaries explicitly.

Agent-readable Embry voice control-plane contract for speaking, listening, cancelling, replaying, and inspecting live voice/chat turns. Use when a project agent needs to control Embry Chatterbox voice, RealtimeSTT listener state, shared Chat UX synchronization, memory/Tau routing, orb state, or session replay through explicit endpoints instead of ad hoc browser scripts.

Best practices for porting features from one codebase, app, harness, TUI, API, or workflow into another without losing destination-specific behavior. Use when users ask to port, clone, migrate, replicate parity, achieve functional equivalence, borrow features, compare source vs destination, or preserve custom work while importing behavior from a reference implementation.

Unified browser automation for AI agents. Uses surf-cli extension when available (full features), falls back to CDP (zero-config). Navigate, read with element refs, click, type, screenshot.

Assess the project to reorganize or deprecate unused/outdated files. Archives large artifacts to 12TB drive, cleans the git workspace, and commits changes.

Persistent browser-oracle tab bindings and directory walk-up registry for WebGPT, Cursor Browser, Gemini, Kimi, and Claude. Binds tab id / viewId + URL once under ~/.pi; maps directories to project names via .ask/browser-oracles.yaml discovered like python-dotenv parent walk-up.

Ingest codebases into /memory for knowledge extraction and CWE scanning. Phase 1 extracts functional knowledge (module docstrings, function signatures, class hierarchies, markdown docs) via Python AST. Phase 2 scans for CWE mappings via /taxonomy. Designed to run nightly via /monitor-codebase.

Standards and workflow for creating root-level PROJECT_STATE.md reports from local receipts, project knowledge, stale/aspirational audits, competitor research, and external reviewer evidence. Use when asked for project state, comprehensive project review, "what works and what is missing", stale code/doc assessment, competitive positioning, or a PROJECT_STATE.md deliverable.

Best practices for leading Ask compete and bakeoff workflows. Use when a user asks for competing models, isolated candidate implementations, winner selection, feature harvesting, approach comparison, model bakeoffs, or a creator competition where $ask should route browser and API handlers through Tau and the project agent must judge results against local evidence.

Best practices for leading multi-model Ask roundtables and Tau-DAG deliberation panels. Use when a user asks for a roundtable, model panel, mixed web/API collaborator discussion, multi-round critique, attributed dissent synthesis, or guidance on turning several model perspectives into executable next steps through $ask.

Red vs Blue team security competition orchestrator. Runs long-running overnight battles with 1000s of interactions, scoring, and insight generation.

Control Chrome browser via CLI for testing, automation, and debugging. Use when the user needs browser automation, screenshots, form filling, page inspection, network/CPU emulation, DevTools streaming, or AI queries via ChatGPT/Gemini/Perplexity/Grok/AI Studio.

Watch any video (URL or local file) with scene-change frame extraction, transcript extraction (native captions or Whisper via scillm), SRT-based emotion/scene analysis, and structured reports. Merges capabilities from ingest-youtube (transcripts) and ingest-movie (scene/emotion analysis) with ffmpeg scene-change detection.

React, Next.js, React Native, and web design best practices from Vercel Engineering. Use when writing, reviewing, or refactoring React/Next.js/React Native components, optimizing performance, auditing UI accessibility, or designing component APIs. Triggers: "review react", "optimize component", "check accessibility", "react best practices", "component architecture", "react native performance".

Continuous SPARTA quality monitor. Runs the 3-tier validation cascade alongside Stage 12 QRA generation — Tier 0 deterministic, Tier 1.5 trained GPT (via /assistant), Tier 2 Brandon (edge cases). Accumulates labels, auto-triggers /create-gpt training at 2000 labels, and tracks convergence.

Deep GitHub repository and code search. For implementation research, uses the sibling brave-search skill to discover candidates, ranks them by requested criteria, stars, and relevance, clones the leaders to /tmp, verifies a detected entry point, then locates and evaluates the requested code.

Browser commands for WebGPT. Agent runs one-liners. All complexity (KDE desktop, CDP stale connections, composer drafts, duplicate tabs, download button clicking) is hidden. Background mode by default — never hijacks the user's mouse or window.

Self-improving PDF extraction convergence loop. Diagnoses extraction failures by computing the delta between S00 estimates and actual extraction, reproduces issues on synthetic PDFs, discovers optimal parameters, and writes fixes back to the extractor pipeline code permanently.

Launch and validate canonical UX Lab adapters and shared UI owned by agent-skills.

Error recovery and anti-patterns for scillm LLM proxy. Load this when scillm calls fail, timeout, or return unexpected results. Covers batch sizing, header requirements, model selection, and debugging workflow.

Use this skill when a project agent needs to spin up, inspect, communicate with, or remove Herdr-managed workstations for long-running multi-agent tasks. It provides a Typer CLI for dynamic Herdr workspaces, provider-specific agent panes, pane-to-pane notifications, semantic role state reporting, and bounded creator/reviewer loops with durable receipts.

Create and update architecture diagrams programmatically in UX Lab's Architecture Editor. Agents define pipeline components, connections, and file attachments — the skill generates proper Excalidraw elements and saves via the Express API. Diagrams are visible at localhost:3002/#architecture and stored in ArangoDB for recall.

Deterministic code review skill with T0 validators (best-practices-*, ruff, compile) and LLM-powered findings (codex/scillm). Scores findings by severity, keeps suggested fixes advisory, and fails closed on provider errors. Structured JSON output. Replaces raw codex exec in orchestrate T2 gate.

Inspect, inventory, repair, normalize, validate, pack, and promote named sprite frames, generated sprite sheets, and contact sheets into exact transparent runtime atlases. Use for missing animation frames, uneven spacing, baked checkerboards, incorrect canvas dimensions, missing alpha, PixiJS manifests, frame-count checks, or sprite-atlas validation.

Compose MIDI arrangements and convert between JSON spec and MIDI files. Two modes: 'compose' uses LLM (via /prompt-lab prompt) to create a full arrangement from annotated lyrics + reference MIDI fragments + heart tags. 'from-spec' and 'to-spec' are mechanical JSON↔MIDI conversion via pretty_midi. The arrangement is the creative heart of the music pipeline — it decides what every instrument plays at every beat.

Research and decide visual casting for persona-dream/movie entities before contact-sheet generation. Use when a story, screenplay, storyboard, or visual_entities.json needs characters, creatures, props, environments, moods, time periods, states, or user-provided reference images turned into a casting_contract.json for downstream contact sheets. Use for "casting agent", "visual casting", "choose references", "research character look", "use these image paths as references", "Brave-search for references", and "produce all contact sheets from story context".

Build or validate provider dry-run packets before any paid media provider call. Use when a storyboard, reference sheets, story contract, or visual packet needs to become a locked provider request with hashes, readiness checks, and explicit paid-call blocking.

Adversarial blind evaluation harness. Generates and runs hidden tests against code written by coding agents. The coding agent never sees the tests — only pass/fail results. Use when /plan creates tasks, or manually with "test-lab run <target>".

Stabilized async task executor. Dispatches plan YAML tasks to local shell commands or code-runner patch runs by default, with broader runners and brittle gates behind explicit opt-in flags. Use when user says "run these tasks", "execute the plan", "orchestrate this".

Create receipt-backed status surfaces and Tau DAG handoff contracts from local progress, gate, and checker receipts. Use when a workflow needs a global progress bar, evidence status board, local readiness surface, or robust Tau connection without claiming provider readiness or live completion.

Shared project knowledge document that human and agent can both read and update. Aggregates state from /project-state, /checkpoint, /memory into a single PROJECT_KNOWLEDGE.md file, while syncing searchable chunks into /memory. Coordinator skill — composes rather than reimplements.

Best practices for Chatterbox or Chatterbox-Turbo voice agents, especially interruptible Embry-style agents that run memory, search, LLM, or other long-running skills concurrently. Use when designing, reviewing, or coding a voice coordinator, async task batch, JSON event stream, cancellable TTS queue, Chatterbox emotion/tag policy, pause policy, barge-in behavior, or spoken progress/preamble strategy.

Best practices for conversational response behavior in voice-first agents: conversation tone, emotional steering, paralinguistic cue injection such as [laughter], wait/delay handling, interruption handling, identity-aware memory grounding, and Chatterbox-ready utterance policy. Use when designing, reviewing, or coding Embry-style chat and voice conversation behavior.

Best practices for authoring, reviewing, and repairing Tau DAG contracts. Use when creating tau.dag_contract.v1 YAML/JSON, choosing project-agent subagent roles, declaring immutable goals, adding skill gates such as best-practices-prompt/react/python, specifying provider/model policy, or diagnosing project-agent DAG failures and tau.dag_error.v1 course-corrections.

MEMORY FIRST - Query memory BEFORE scanning any codebase. Use when encountering ANY problem, error, or task. Call "recall" FIRST, then scan codebase only if nothing found. Triggers: "check memory", "recall", "have we seen", "remember how".