01 — SkillOnOff

What it shows#

The smallest possible "skill" demo: send the same prompt to a local Ollama model twice, once with a vanilla system prompt and once with a short skill file prepended to the system prompt, then print both answers side-by-side. The lift from a 5-line skill is usually obvious at a glance — that's the entire point. No agent framework, no orchestration, just HttpClient and a Markdown file.

Prerequisites#

.NET 10 SDK
Ollama installed and running (ollama serve)
A model pulled locally:
```
ollama pull llama3.2
```
Optional: override the model with $env:OLLAMA_MODEL = "qwen2.5:3b".

Run#

From this folder:

dotnet run -- "Write a short bio for a senior developer." concise-tone

Both arguments are optional. Defaults: prompt = "Write a short bio for a senior developer.", skill = concise-tone.

Try the second built-in skill:

dotnet run -- "Explain garbage collection in one paragraph." pirate-voice

Sample output#

Model:  llama3.2
Skill:  concise-tone
Prompt: Write a short bio for a senior developer.

Calling Ollama twice (raw + skill)...

─── ─ RAW ─────────────────────────────────────  ─── ─ WITH skill: concise-tone ─────────────────
Jane Doe is a passionate and highly experienced  - 12+ years building distributed systems
senior software developer with over twelve       - Specializes in C#/.NET, Go, Kubernetes
years of professional experience building        - Led platform team of 8 engineers
distributed systems at scale. She specializes    - Open-source maintainer, frequent speaker
in C#, .NET, Go, and Kubernetes, and has led     - Mentors junior devs across 3 timezones
platform teams of up to eight engineers across   - Coffee, vinyl, mountain biking
multiple time zones. In her spare time she
enjoys mountain biking and curating her vinyl
collection.

(example output — your model will phrase things differently)

How it works#

Loads skills/{name}.skill.md, strips the YAML frontmatter, keeps the body as a system-prompt fragment.
Fires two POST http://localhost:11434/api/chat calls in parallel via Task.WhenAll:
- Raw: system = "You are a helpful assistant."
- Skill: system = "You are a helpful assistant.\n\n" + skillBody
Both calls use stream: false so we get one tidy JSON blob each.
A tiny word-wrap helper prints both responses in two ~50-column columns so the lift is obvious live.

Try this#

Built-in skills:

File	Effect
`skills/concise-tone.skill.md`	Terse bullets, no adjectives, hard cap on length.
`skills/pirate-voice.skill.md`	Rewrites every reply as an 18th-century pirate.

Drop your own *.skill.md into the skills/ folder — frontmatter (name, description) plus a body — then pass the filename (without the extension) as the second argument:

dotnet run -- "Summarize the CAP theorem." my-custom-skill

That's the whole "skills" pillar in ~150 lines of C#.