tiziano-ai/charter-skills

3D art for rendered stills and films or interactive scenes: create or repair models, scenes and their motion (Blender, game and web engines), or keep their look at lower render cost. Art direction, form, materials and light; not gameplay rules, generated image/video or engine setup. A rendered film also needs visual asset production for its finished file.

Compose established meaning for its receiver, purpose and medium, reducing avoidable interpretation effort while preserving precision, relationships and expressive qualities.

Learn from recorded agent sessions and transcripts: what went wrong and why, what worked, repeated corrections and changes of direction; return ready changes for whoever can make them. Not a recap or catch-up, live recovery or automatic instruction rewriting.

Design CLIs, MCP servers and plugins or extensions that agents call: operations, identities, results, errors, host discovery, versions and what survives disable or uninstall, so callers act without reading source. Operating an existing tool follows its guide.

AI evaluation (evals) for model outputs, prompts and agents in products, when observed quality decides adoption, repair or a model change: expert-chosen hard cases, error analysis before metrics, validated LLM judges, cost.

Design native Apple interfaces when scene, input or platform choices shape the workflow; create coherent navigation, commands, adaptation and accessibility across Apple platforms and framework seams. Apple-styled browser pages remain web-interface work.

Architecture simplification before building or refactoring: use suitable native, standard or upstream primitives to remove unnecessary responsibilities and custom layers while preserving required capability.

Reorganize file collections when retrieval or interpretation breaks down; choose taxonomy and names, plan recoverable moves and keep provenance, bundles and versions. Deleting identical copies is a separate job.

Find why a setting does not take effect: the value in the file is not what the running process uses. Trace its reader, precedence, launch environment and load time, then fix the input that wins. Not machine setup or release delivery.

Design or repair retrieval and context assembly for LLM products: carry the right sources, relationships and history into usable answers. Choose parsing, search, ranking, filters and context budget from representative questions; not general prompt writing or runtime session recovery.

Design source changes that span schemas, producers and consumers; carry one coherent meaning through every behavior-bearing layer and generator, retiring competing paths while preserving real external compatibility. A clear local edit needs no cross-layer design.

Resolve external facts whose freshness or applicability is uncertain; reconcile source authority, scope and effective dates to answer the actual question with supported conclusions and consequential uncertainty.

Decision design for consequential choices involving uncertainty, competing preferences or shared commitments: integrate existing contributions, compare feasible options and recommend an actionable choice. Market bets are product strategy.

Create or repair documents when destination fidelity is at stake; preserve required layout, structure, calculation, extraction and editability through rendering, export, import or save. Compose with format-specific authoring tools rather than replacing content expertise.

Design the substantive work of a new or incomplete product: turn inputs into usable findings, decisions or actions and assign each transformation. Not ordinary implementation of a clear function, market strategy or the journey through the product.

Game design for creating or improving how a game plays: connect rules, information, controls, pacing and progression to the intended player experience. Not engine setup, interface styling or technical QA alone.

Instruction system design when agent guidance conflicts, misfires or has unclear authority, or after a model or harness upgrade: choose what to teach, where each rule lives and how layers compose. Not a wording comparison or retrieval repair.

Diagnose abnormal CPU, GPU or ANE, unified-memory, swap, thermal, energy, or I/O use on macOS by joining Activity Monitor, system pressure, process and model identity, stacks, logs, and controller state. Use for hot daemons or apps, unexplained memory pressure, GPU load, or loaded local models in MLX, Ollama, LM Studio, llama.cpp, and similar runtimes; not for source-level profiling after one application and code path are already known.

Mac disk cleanup: remove rebuildable caches, dependencies and build outputs without breaking active or source-linked consumers; reclaim already-free container storage. Not duplicates, models, histories, installed tools or data retirement.

Design or improve product journeys from a user's need to a retained result: first success, comprehension, trust, recovery, control and desire to return. Market positioning, substantive product functions and platform implementation have separate owners.

Choose product strategy when market, advantage or growth bets are open; connect whom to serve, why they switch, distribution and economics into a defensible direction and sequence of bets. Improving an already-chosen user journey is product-experience work.

Prompt design for a system prompt, task prompt or agent role: write or improve it, adapt it to another model or provider, or choose between complete candidates, controlling judge bias and naming the cheapest trial that would settle a close call.

Design capability support across providers, models or backends; decide selection, adaptation, degradation and fallback across transports and platforms without losing the user's requested behavior or constraints. Account setup and runtime operations belong to the owning integration.

Release, deploy or install an exact change so the installed or hosted consumer actually runs it; resolve artifact and target identity, runtime selection, compatibility, cutover and recovery. Source readiness alone is not delivery.

Assess or catch up on a repository that is new to you or has changed: trace a real product path to what it supports, where behavior is owned and which gap decides the first useful work. Deploying or installing it is release engineering.

Session coordination when agent sessions or tasks collide or depend on each other, in one runtime or across harnesses: resolve competing writes, implementations or contribution handoffs so shared work has one integration owner. Independent parallel work and ordinary delegation need none.

Status across the agent sessions or tasks this runtime can list: explain from current evidence what is usable, running or waiting and where a decision matters, without continuing the work. Inspecting one session, coordination and stopping sessions are separate jobs.

Session wind-down before a restart, pause, handoff or account switch, across the agent sessions or tasks this runtime can list: let in-flight work finish, start nothing new, keep the way back and report what is ready. Does not restart, resume or move them.

Skill package design when creating a SKILL.md skill or fixing one that misroutes, goes unused or underperforms: decide what belongs in it, write a truncation-proof card, put decisive rules in the body. Native creators scaffold.

Design or repair UI and CLI status when shown state can mislead action; align progress, freshness, outcomes, diagnostics, history and offered controls with real state. Recovering the stuck work is workflow recovery.

Repair streaming protocols when event semantics corrupt outcomes; fix framing, ordering, partial output, callbacks, cancellation, reconnect, replay or competing terminal signals at the responsible seam. Ordinary command waiting and recovery of one durable job are separate work.

Diagnose unexplained slowness or recurring, cross-boundary failures; explain the mechanism in interacting owners, state, queues, feedback or incentives and choose the nearest sufficient intervention. Not for an already-localized bug.

Software tests: adding tests, fixing a flaky one, or finding why a test passes or fails; choose expected values, boundaries and mocks that catch the real bug, and drop tests that protect nothing.

Produce image or video assets for a concrete placement, including concept takes while a look is open, texture imagery and the finished file of a film rendered from a 3D scene (3D art builds and renders it); direct, select, repair and derive coherent variants with useful provenance and integration fitness. Whole-interface design is separate.

Design or repair websites and web apps: landing pages, portfolios, editorial pages and working tools, with subject-led visual character, responsive composition, navigation, accessible controls and coherent asynchronous state. Native Apple interaction and standalone visual assets are separate jobs.

Work discovery when a need is hard to articulate: discover the intended result and scope through adaptive interviewing, evidence and exploration, and carry that understanding into the result. Clear requests need ordinary execution.

Recover interrupted workflows with uncertain surviving effects; reconcile operation state and durable output to find the first safe continuation without duplicating work. Healthy processes, ordinary waits and broken protocol design need different treatment.