Skip to content

ericrisco/rsc-harness / skills/ollama

ollama

Verified

Use when running open-weight LLMs locally with Ollama — pulling and tagging models, calling the local API, picking a quantization or GGUF, writing Modelfiles, and sizing VRAM and RAM for the machine at hand. NOT remote or managed GPU serving and autoscaling (that is `runpod`), NOT downloading raw weights or datasets (that is `huggingface`), NOT retrieval pipeline design (that is `rag`).

FileKindSize
SKILL.mdskill md11.2 KB
evals/README.mdmarkdown714 B
evals/cases.yamlconfig3.0 KB
references/api.mdmarkdown5.1 KB
references/hardware-sizing.mdmarkdown4.0 KB
scripts/verify.shexecutablescript4.8 KB

evals/cases.yaml

View on GitHub
skill: ollama

should_trigger:
  - prompt: "How do I run Llama 3.3 70B locally with Ollama?"
    why: Core local-serve request — pull/run a named model on this machine.
  - prompt: "My Ollama model keeps OOMing once the conversation gets long — how do I fix it?"
    why: Non-obvious — the cause is KV-cache growth with context; fix is num_ctx cap / OLLAMA_KV_CACHE_TYPE / smaller model, all owned here.
  - prompt: "Write a Modelfile that gives qwen3 an 8k context and a terse system prompt."
    why: Modelfile authoring (FROM + SYSTEM + PARAMETER num_ctx) plus ollama create.
  - prompt: "Point my OpenAI Python SDK at a local model instead of api.openai.com."
    why: OpenAI-compat layer at localhost:11434/v1 with a dummy key — a non-obvious 'it just works' path.
  - prompt: "Quina quantització (Q4 o Q8) i quin model caben a la meva RTX 3060 de 12GB?"
    why: Catalan; quant choice + VRAM sizing for a specific GPU, the heart of the skill.
  - prompt: "Convert this GGUF file into an Ollama model and quantize it to Q4_K_M."
    why: FROM ./model.gguf import plus ollama create --quantize q4_K_M.
  - prompt: "ejecutar un LLM en local sin API key, totalmente offline"
    why: Spanish; the local/offline/no-key framing that defines Ollama's niche.

should_not_trigger:
  - prompt: "Deploy this model on a serverless GPU endpoint that autoscales with traffic."
    route_to: modal
    why: Managed/serverless autoscaled serving — Ollama is single-box, no autoscale.
  - prompt: "Rent a GPU in the cloud to serve a 70B model behind an HTTP endpoint."
    route_to: runpod
    why: Remote/managed GPU rental and serving, not local inference.
  - prompt: "Download the Llama weights and a fine-tuning dataset from the Hub with the hf CLI."
    route_to: huggingface
    why: Weights/dataset/repo management — Ollama imports a result, it doesn't manage the Hub.
  - prompt: "Build a retrieval pipeline that chunks PDFs and reranks before calling the model."
    route_to: rag
    why: Retrieval/chunking/reranking design; the model is the downstream consumer.
  - prompt: "Help me write a better system prompt for tone and refusal behavior."
    route_to: prompt-engineering
    why: Prompt content authoring, not model serving or hardware.

capability:
  - scenario: "I have an RTX 4070 with 12GB VRAM. I want a coding assistant running locally via Ollama, called from a Python app, with a custom 16k-context system prompt."
    must_include:
      - Recommends a model size + quant that fits ~12 GB (e.g. 7B–14B at Q4_K_M) with the sizing reasoning (weights ≈ params × bytes/param × 1.2).
      - Warns that a 16k context adds KV cache on top of weights; suggests PARAMETER num_ctx 16384 and OLLAMA_KV_CACHE_TYPE=q8_0 if tight.
      - Gives a Modelfile (FROM + SYSTEM + PARAMETER num_ctx 16384) and the ollama create command.
      - Shows calling it from Python — native /api/chat or the OpenAI-compat /v1 with base_url http://localhost:11434/v1 and a dummy key.
      - Names the boundary — if it doesn't fit or needs scale, move to runpod/modal.
writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deploymentStars: 322kUpdated Claude CodeCodexVerifiedInstall
xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy.Stars: 119kUpdated Claude CodeCodexVerifiedInstall
childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release.Stars: 118kUpdated Claude CodeCodexVerifiedInstall