AI & ML

catcorner22 Bundle

Use to select models to run locally with llama.cpp and GGUF on CPU, Mac Metal, CUDA, or ROCm. Covers finding GGUFs, quant selection, running servers, exact GGUF file lookup, conversion, and OpenAI-compatible local serving.

catcorner22 Bundle

Use this skill when the user wants to build tool/scripts or achieve a task where using data from the Hugging Face API would help. This is especially useful when chaining or combining API calls or the task will be repeated/automated. This Skill creates a reusable script to fetch, enrich or process data.

catcorner22 Skill

Pause a LangGraph graph for a human: `interrupt()`, `Command(resume=...)`, approval and validation workflows, resuming multiple concurrent interrupts by id, keeping pre-interrupt side effects idempotent, and the 4-tier error-handling strategy. Use when a LangGraph run must stop for human input or approval. Scope boundary: `HumanInTheLoopMiddleware` on a `create_agent` agent belongs to `langchain-middleware`, approval gates inside the `deepagents` harness to `deep-agents-orchestration`, and the checkpointer that makes a resume possible to `langgraph-persistence`.

catcorner22 Skill

Independent agents score rubric dimensions with written ballots; synthesis agent explains weighting. Use for high-stakes evaluation. Scope boundary: counterpoint generation → `counterpoint-perspectives`; wine blend → `wine-blending-fusion`.

catcorner22 Skill

Dual retrieval: embedding similarity PLUS taxonomic adjacency in task ontology. Use when related concepts use different wording. Scope boundary: triangulation for facts → `survey-triangulation`.

catcorner22 Skill

Token- or block-level provenance: which inputs, tools, prompt sections, and turns produced each output segment. Use when debugging AI failures or audit requirements. Scope boundary: claim-level sources → `journalistic-attribution`; orient logging → `ooda-adaptive-context`.

catcorner22 Skill

Borrow a foreign genre skeleton (contract, tech spec) and plant the current content into it. Use when the native outline is weak. Scope boundary — graft a specialist passage → `orchard-graft-transfer`; multi-model blend → `wine-blending-fusion`.

catcorner22 Skill

Calibrate generation parameters as just-intonation ratios by task type (factual, creative, code) instead of one global temperature. Use when the same model is too loose on facts or too stiff on drafts. Scope boundary — prompt wording itself → `prompt-optimizer`; compute budget class → `emergency-triage-compute`.

catcorner22 Bundle

Run evaluations for Hugging Face Hub models using inspect-ai and lighteval on local hardware. Use for backend selection, local GPU evals, and choosing between vLLM / Transformers / accelerate. Not for HF Jobs orchestration, model-card PRs, .eval_results publication, or community-evals automation.

catcorner22 Bundle

Publish and manage research papers on Hugging Face Hub. Use to create markdown research-article drafts, check or GET-index a paper page, link arXiv IDs into model/dataset/Space READMEs, and generate citations. Claim authorship, POST-index, and paper search belong in the huggingface-papers skill.

catcorner22 Bundle

Train or fine-tune sentence-transformers models across `SentenceTransformer` (bi-encoder; dense or static embedding model; for retrieval, similarity, clustering, classification, paraphrase mining, dedup, multimodal), `CrossEncoder` (reranker; pair scoring for two-stage retrieval / pair classification), and `SparseEncoder` (SPLADE, sparse embedding model; for learned-sparse retrieval). Covers loss selection, hard-negative mining, evaluators, distillation, LoRA, Matryoshka, and Hugging Face Hub publishing. Use for any sentence-transformers training task.

catcorner22 Bundle

Build agents in Python with Pydantic AI — tools, capabilities (including on-demand loading), structured output, streaming, testing, and multi-agent patterns. Use when the user names Pydantic AI, when code imports `pydantic_ai` or uses its `Agent`/`RunContext`/`Tool` classes, or when a Python AI agent needs tools/capabilities, deferred capability loading, streamed output, YAML-defined agents, or agent tests. Scope boundary: this skill covers Pydantic AI in Python only — TypeScript/JavaScript agent work and framework-unspecified agent-building requests belong to `build-agents`/`eve`, and the `pydantic` validation library on its own is out of scope.

catcorner22 Skill

Access and test Vercel deployments protected by Vercel Authentication, SSO, or Deployment Protection. Use when curl, agent-browser, Playwright, or another automated request reaches a Vercel login or protection page; when a protected preview or production URL returns 401 or 403; when TRUSTED_SOURCES_ENVIRONMENT_MISMATCH appears; or when choosing between `vercel curl` and the `x-vercel-trusted-oidc-idp-token` header.

catcorner22 Bundle

Create LangSmith online evaluators one at a time: inspect recent traces in a tracing project, interview the user, propose grounded criteria, then build, test, attach via a run rule, and verify a single LLM-as-judge or code evaluator per pass. Use for evaluation attached to live LangSmith traces. Scope boundary: offline model benchmarking and leaderboards belong to `huggingface-community-evals`.

jacob-balslev Bundle

Use when deciding whether to run a task in the Codex CLI agent harness (which drives a frontier GPT model), when scoping work to its native capabilities (resumable `codex exec resume` sessions, non-interactive `exec` stdout-piping, in-process dispatch, `/permissions` sandbox modes, MCP, on-demand subagents, the `/review` code-review agent), or when choosing Codex versus the Claude Code harness for a piece of work — and when avoiding its known failure modes (non-autonomous subagents, Full-Access network reach, cold-one-shot context loss). Do NOT use for routing a task to the GPT MODEL versus Claude (use `gpt-5-5`), for the Claude Code harness decision (use `claude-code`), or for designing a generic resumable agent loop (use `autonomous-loop-patterns`).

jacob-balslev Bundle

Use when deciding whether to route a task to OpenAI's GPT-5.5 frontier model versus Claude Opus or Sonnet — picking the model lane for infrastructure scripts, CI pipelines, concrete implementation, analytical code review, security review, or CLI/terminal-heavy work, and weighing GPT-5.5's context window, pricing, and per-benchmark strengths against the Claude tiers for the same task. Covers the decision-useful capability and pricing facts, and the boundary against the Claude routing skills. Do NOT use for running the GPT model through a harness (use `codex`), for choosing among the Claude tiers themselves (use `claude-opus` / `claude-sonnet`), or for routing among local skills at request time (use `skill-router`).

jacob-balslev Bundle

Use when deciding whether to run a task on the OpenCode agent runtime (the open-source, provider-agnostic coding agent — usable as a terminal TUI, a non-interactive `opencode run` CLI, an IDE bridge via ACP, a desktop app (beta), a web UI, or a headless `opencode serve` server), how to invoke it non-interactively, how to pick a `provider/model` string, or how OpenCode differs from Claude Code and Codex. Covers its 75+-provider model routing, the OpenCode Zen curated/free model service and the separate OpenCode Go subscription lane, model-selection and config-source (8-layer) precedence, the `--format json` JSONL event stream, the full CLI command surface (serve/web/attach/acp/session/stats/export/import/github/pr/plugin/mcp/db/debug), built-in build/plan agents and LSP code intelligence, permissions vs provider policies, native Agent Skills / MCP / references / commands extension layers, local models, and scripting/automation. Do NOT use when choosing WHICH free model to route specific work to (use `opencode-

jacob-balslev Bundle

Use when working with Etsy listings — Etsy's own title/tag/category/attribute/description/photo/video model and two-phase search (query matching → ranking on relevancy, listing engagement, recency, shipping price, customer-and-market experience, and personalization); the 13-tag and 20-character-per-tag limits and 140-character title cap; Etsy's 2026 natural-readable-title guidance; the description as an indexed/semantic and Google-snippet surface; free-shipping/shipping-price visibility thresholds; listing-video specs; the first-party Search Visibility diagnostic; the Etsy Open API listing surface; and editing Etsy listings managed by an upstream POD platform (Printify/Printful) including the publish/lock lifecycle, selective publishing, and publish-failure diagnostics. Do NOT use for generic Shopify storefront work, non-Etsy marketplaces, deep keyword-research methodology, the POD product/blueprint/variant model itself, or print-file/artwork generation.

jacob-balslev Bundle

Use when deciding whether to route a task to Google's Gemini Pro frontier model (current Gemini 3.1 Pro generation, moving to Gemini 3.5 Pro) instead of Claude Opus or GPT-5 — especially for very-large-context reasoning (1M-token input), whole-codebase or long-document analysis, native multimodal input (audio/video/PDF/image in one call), or capability-per-dollar at the frontier. Covers the context window, the 200K context-tier pricing cliff, multimodal support, and the per-lane comparison against Claude Opus / GPT-5. Do NOT use for choosing the cheap/fast tier (use `gemini-flash`), for general agent-system architecture (use `agent-engineering`), or for dispatching among local skills (use `skill-router`).

jacob-balslev Bundle

Use when deciding whether to run a task in the Claude Code agent harness, when scoping work to its native capabilities (skills, hooks, subagents, MCP, plan mode, 1M-context Opus, background tasks, slash commands), or when choosing Claude Code versus a different harness (Codex, OpenCode, Copilot) for a given piece of work. Covers what the harness is good at, the extensibility stack and when each layer earns its keep, and the decision boundaries against rival harnesses. Do NOT use for writing Anthropic SDK / API code (use `claude-api`), for choosing which Claude MODEL to route a task to versus GPT (use `gpt-5-5`), or for designing a generic autonomous agent loop (use `autonomous-loop-patterns`).

jacob-balslev Bundle

Use when deciding whether to route a task to Anthropic's frontier reasoning tier (Claude Opus) — architecture, multi-file synthesis, hard debugging, security reasoning, long-horizon agentic planning — and when NOT to (mechanical work belongs on a cheaper tier). Covers the Opus capability profile: 1M-token context, the cost/latency premium, adaptive thinking, the Opus-only effort ceiling (xhigh/max), task budgets, prompt-caching minimums, and high-resolution vision. Do NOT use for picking the balanced implementation tier (use claude-sonnet), the fast/cheap tier (use claude-haiku), designing the loop the model runs inside (use autonomous-loop-patterns), or for Claude API request syntax (read the claude-api reference).

jacob-balslev Bundle

Use when deciding whether to route a task to the fast/cheap tier (Claude Haiku) — transcription, polling, format conversion, structured-output slot-filling, small-diff review, high-volume low-latency work — and where the boundary is that should escalate to Sonnet/Opus. Covers the cost/latency floor, the 200K context ceiling (vs the upper tiers' 1M), the absence of the effort knob, the separate rate-limit pool, and when to drop below Haiku to a script. Do NOT use for ordinary multi-step feature work (use claude-sonnet), the hardest reasoning (use claude-opus), loop design (use autonomous-loop-patterns), or Claude API request syntax (read the claude-api reference).

jacob-balslev Bundle

Use when deciding whether to route a task to the balanced implementation tier (Claude Sonnet) — feature work, bug fixes, test writing, multi-step code — as the default lane that is cheaper/faster than the frontier tier and more capable than the fast tier. Covers the cost/quality tradeoff vs Opus and Haiku, the shared 1M context window, effort behavior, and the 1M-context subscription billing caveat. Do NOT use for the hardest reasoning/architecture/security work (use claude-opus), high-volume mechanical or low-latency work (use claude-haiku), loop design (use autonomous-loop-patterns), or Claude API request syntax (read the claude-api reference).

jacob-balslev Bundle

Use when designing or auditing the multi-graph context architecture of an AI-coding workspace: skill graph, document routing graph, memory index, script registry, and the cross-graph edges between them. Covers edge typing, orphan detection, connectivity health, deterministic graph synthesis signals, change-propagation checks, and drift or hub-and-spoke anti-patterns. Do NOT use for authoring one SKILL.md (use `skill-scaffold`), validating one skill (use `skill-infrastructure`), live routing decisions (use `skill-router`), context-window budgeting (use `context-window`), or session load/drop choices (use `context-management`).

jacob-balslev Bundle

Use when working with Printify — the print-on-demand REST API, catalog model (blueprints, print providers, variants, print areas), product creation and publish lifecycle to connected channels, order routing, shipping cost queries, and HMAC SHA-256 webhook verification. Do NOT use for non-Printify POD vendors, generic Shopify storefront work, or print-file (artwork) generation.

jacob-balslev Bundle

Use when allocating context-window budget across system, skill-injection, working, and output zones; monitoring context health; deciding when to compact; preserving state before compaction; recovering after compaction; or choosing strategies for 1M, 200K, or 128K context windows. Covers zone budgets, practical model-budget tables, the 80% compaction rule, pre/post-compact protocols, persistence hierarchy, operation token costs, and token-reduction techniques. Do NOT use for deciding what information belongs in the working set (use `context-management`), prompt design (use `prompt-craft`), graph architecture (use `context-graph`), or memory curation.

jacob-balslev Bundle

Use when deciding whether to spend GitHub Copilot's metered budget on a task (its premium-request / post-June-2026 AI-credit model, where 1 credit = $0.01 and cost = tokens × per-model rate), what Copilot is good at vs expensive at, which plan allowance (Pro 300 / Pro+ 1500, no rollover) applies, and when a cheaper or free lane should take the work instead. Covers the June 1 2026 shift from premium-request multipliers to usage-based token billing, the always-free completions/next-edit surface, and the IDE-native frontier-model lane. Do NOT use for choosing or operating the OpenCode runtime (use `opencode`), for picking a specific free model (use `opencode-free-models`), or for authoring an agent loop (use `autonomous-loop-patterns`).

jacob-balslev Bundle

Use when creating a new SKILL.md from scratch, restructuring a draft before it becomes a stable skill, or teaching another author the canonical Skill Metadata Protocol frontmatter, body, and audit-state.json sidecar structure. Covers flat schema-conformant frontmatter, the sidecar split, v8 classification, body layout by skill intent, semantic-layer discipline (description vs activation vs Coverage), teaching-layer mechanics (TEMPLATE NOTE blockquotes), native skill-creator handoff, public/private safety, and routing-eval honesty. Do NOT use when modifying an already-written skill (edit it directly), writing general technical documentation, routing an existing request across skills, or fixing malformed library health at scale (use `skill-infrastructure`).

jacob-balslev Bundle

Use when scoring a completed agent task, implementation, document, skill upgrade, or other deliverable against the original request, acceptance criteria, verification evidence, quality rubric, and residual risks before calling it done. Covers skeptical critic review, 1-5 scoring, score ceilings, evidence sufficiency, finding/action capture, and the evaluation-revision loop. Do NOT use for designing eval datasets or graders (use eval-driven-development), line-by-line diff review (use code-review), choosing test levels (use testing-strategy), or designing the overall process and gates before work starts (use methodology).

jacob-balslev Bundle

Use when turning raw qualitative research output (interview transcripts, field notes, diary entries, observation logs) into themes, patterns, insight statements, and design-ready artifacts via affinity mapping, empathy maps, and jobs-to-be-done framing. Do NOT use for collecting new research, quantitative analysis, statistical inference, or summarizing a single document — synthesis specifically operates on a corpus of qualitative evidence.

jacob-balslev Bundle

Use when writing, tightening, evaluating, or repairing an LLM prompt or reusable prompt template for completion, agent dispatch, grading, structured extraction, tool use, or prompt-engineered workflows. Covers instruction hierarchy, message roles, context placement, few-shot examples, structured output, positive constraints, reasoning guidance, prompt-injection resistance, provider differences, and eval-driven iteration. Do NOT use for whole context-system design (use context-engineering), eval dataset or grader design (use eval-driven-development), reviewing generated code (use code-review), authoring SKILL.md files (use skill-scaffold), choosing which skill or agent should activate (use skill-router), or root-causing a deployed failure after outputs already exist (use debugging).

jacob-balslev Bundle

Use when deciding what to load into an active agent session, recovering from context drift, preparing compaction or restart, distilling raw inputs into a working summary, or writing a handoff another agent can resume quickly. Covers intake triage, the active-context contract, the six-step context-management loop, working-set shaping, evidence capsules and observation masking, just-in-time handle-before-payload loading, prompt-caching alignment, context-rot defenses, drift signals and reset, the runtime-primitives boundary, compaction-ready handoffs, and selective rebuild after context loss. Do NOT use for token math (use `context-window`), prompt wording (use `prompt-craft`), persistent memory curation, or multi-graph context architecture (use `context-graph`).

jacob-balslev Bundle

Use when condensing prose while preserving meaning: session findings, wrap reports, research briefs, executive summaries, TLDRs, agent handoffs, progressive summaries, audit summaries, and long-document distillation. Covers extractive vs abstractive summarization, what to keep vs drop, evidence preservation, summary levels, handoff summaries, and audit-report condensation without hiding findings. Do NOT use for byte/data compression algorithms (use `compression`), context-window budget math or compaction triggers (use `context-window`), working-set selection (use `context-management`), prose tone repair (use `writing-humanizer`), or quality scoring (use `evaluation`).

jacob-balslev Bundle

Use when designing what information reaches an LLM agent before it reasons — system prompt, persistent memory, always-loaded rules, injected skills, and the user prompt — or when diagnosing why an agent produced a wrong answer despite a clear instruction. Covers the four context failure modes (missing, stale, wrong, overwhelming), the five-layer context stack, four context quality metrics (injection precision and recall, utilization, freshness), the Frequent Intentional Compaction (FIC) protocol, subagent delegation for context-heavy work, and the failure-mode decision tree. Do NOT use for prompt wording (use `prompt-craft`), authoring a new SKILL.md (use `skill-scaffold`), or deciding which skill the router activates for a given query (use `skill-router`).

jacob-balslev Bundle

Use when reasoning about the protocol-level cycle by which a language model uses external tools: declaration, request, execution, continuation, model-visible transcript or provider-resumed state, ID pairing, tool-result formatting, vendor encodings (Anthropic Messages, OpenAI Responses and Chat Completions, MCP, Gemini generateContent and Interactions), client tools vs hosted/server tools, strict schemas and grammar-constrained custom tools, streaming arguments, tool search/deferred loading, programmatic tool calling, computer-use/browser-control screenshot-action loops, parallel vs sequential calls, error handling, and the separation between model intent and runtime execution. Do NOT use for deciding when or how many tools to call (use tool-call-strategy), multi-agent architecture (use agent-engineering), prompt wording (use prompt-craft), or eval design for tool-use behavior (use eval-driven-development).

jacob-balslev Bundle

Use when reviewing a pull request, diff, or proposed code change for correctness, clarity, security, performance, maintainability, test evidence, and conformance to project conventions — whether the author is a human, an AI agent, or a peer. Covers pre-review fact-gathering (including verifying an AI-written PR summary against the diff), the read-order strategy (tests first, then implementation, then call sites and blast radius), the size/attention budget, the severity-grading rubric mapped to the Conventional Comments standard, comment-phrasing discipline, reviewer qualification, the rule that diff content is evidence not instructions, the no-rubber-stamp rule for AI-generated diffs, and the explicit approve/request-changes/close merge decision. Do NOT use for AUTHORING the code (use `refactor` for behaviour-preserving changes or `skill-scaffold` for new skills), for chasing a known bug after merge (use `debugging`), for security-only audits (use `owasp-security`), or for explaining a patch without a merge v