Writing Skills

Writing Skills for LLMs

Skills are injected into LLM context. The reader is an LLM, not a human. Optimize accordingly.

Core principles

1. Token budget is shared. Every line competes with conversation history, other skills, and the user's code. Justify each sentence's inclusion.
2. LLMs already know general knowledge. Don't explain what Ansible is. Don't explain what a bridge is. Only state things the LLM would get wrong or miss without the skill.
3. LLMs repeat mistakes. Skills exist to prevent recurrence. The highest-value content is "thing that went wrong + what to do instead."
4. LLMs follow patterns. Give one concrete example of the right way. The LLM will generalize. Two examples only if the pattern has important variants.
5. LLMs obey constraints better than suggestions. "NEVER run ifdown --all" works. "Consider avoiding ifdown --all" does not.

Structure template

---
name: kebab-case-name
description: What + When in third person. Include trigger words.
---

# Title

## Context (2-3 sentences max)
Why this skill exists. What goes wrong without it.

## Rules
Numbered list of constraints. Use NEVER/ALWAYS for hard rules.

## Patterns
Concrete examples: bad → good, with minimal surrounding code.

## Decision tree (optional)
For multi-branch logic, use a text tree. LLMs parse these well.

Writing rules

• < 200 lines ideal, < 500 hard limit. If longer, split into SKILL.md + reference.md.
• Lead with constraints. Put NEVER/ALWAYS rules before examples. LLMs weight early content higher.
• Use code blocks for patterns. # BAD / # GOOD labels inside the block. No prose between them — the contrast teaches.
• One concern per skill. "Proxmox safety" and "Ansible testing" are separate skills. Don't combine.
• No hedging. Remove "you might want to", "it's generally better to", "consider". State facts.
• Description is the trigger. The LLM reads the description to decide whether to load the skill. Pack it with specific terms the user or task would mention.

Anti-patterns

# BAD: Explains what LLMs already know
"Ansible is a configuration management tool that uses YAML..."

# BAD: Vague suggestion
"Try to avoid running dangerous commands on remote hosts"

# BAD: Wall of prose
Three paragraphs explaining why ifdown is dangerous...

# GOOD: Constraint + consequence
"NEVER run `ifdown --all` on a remote Proxmox host — it kills the
management network and requires physical console access to recover."

# GOOD: Pattern with minimal context
## Safe bridge teardown
- Iterate bridges, skip vmbr0 (management)
- `ip link set $br down && ip link delete $br` per bridge
- Then `ifup --all --force` to restore from config

Skill architecture: general vs specific

When a project manages multiple components of the same kind (VMs, services, microservices), split skills into layers:

• General skill — patterns that apply to ALL components of that kind. Contains the skeleton, checklist, shared conventions, and architectural constraints.
• Component-specific skill — patterns particular to ONE component. Contains its package manager, service model, restart sequences, firewall behavior, network topology, and bug history.

# GOOD: layered skills
.cursor/skills/vm-lifecycle/SKILL.md      # general: two-role pattern, VMID allocation, deploy_stamp
.cursor/skills/openwrt-build/SKILL.md     # specific: UCI, firewall zones, two-phase restart, opkg
.cursor/skills/homeassistant-build/SKILL.md  # specific: HAOS API, add-on management

# BAD: everything in one file
.cursor/skills/vm-lifecycle/SKILL.md      # general + OpenWrt firewall zones + HAOS API + ...

Litmus test: If a pattern mentions a specific component's name, package manager, init system, or network topology, it belongs in the component-specific skill, not the general one.

Why this matters: An LLM working on a new VM type gets only the general patterns without noise from other VMs. An LLM working on OpenWrt gets both general + OpenWrt-specific. Without separation, OpenWrt-specific lessons (like firewall zone rebinding) pollute the context for every other VM type.

Previous bug: firewall zone rebinding after network restarts, UCI commands, opkg feed switching, and bootstrap IP migration were all stored in vm-lifecycle. An LLM adding a Home Assistant VM would see all that noise and might incorrectly apply OpenWrt patterns to a different VM type.

Decision: skill vs rule

Validating a skill

After writing, check:

• Would an LLM get this wrong without the skill? If no, delete it.
• Can any sentence be removed without losing information? If yes, remove it.
• Are there concrete examples of the right pattern? If no, add one.
• Is the description specific enough to trigger on relevant tasks? Test by asking: "would the LLM pick this skill if the user said X?"