AI agents are only as good as the context you give them. To shape how your agent behaves and what it knows how to do, there are two concepts at your disposal. Rules and skills allow you to spend less time repeating yourself and more time getting better results.
What You Should Know
After working with an AI agent for a period of time, you may have noticed it can be inconsistent. It might nail a task one session and completely miss the mark for the same task in the next. More often than not, the issue isn’t the model or even hallucinations. Maybe your agent just doesn’t know your expectations. Rules and skills are two ways to solve that problem and give agents the context they need to be successful.
Rules
A rule is a persistent instruction your agent follows every time it works, across every session. Think of rules as the standing agreements between you and your agent. They define how it should communicate, what conventions it should follow, what it should never do without asking first.
Some examples:
“This project uses 2-space indentation and single quotes.”
“Flag accessibility concerns before suggesting visual changes.”
“All CTAs should follow the voice and tone guide when drafting or editing copy.”
“Summarize decisions and action items at the end of every working session.”
Rules are global and ongoing. They remove the need to re-explain your preferences at the start of every conversation. Tools like Windsurf and Cursor have built-in recognition system (typically a hidden editor directory, like .cursor) where you can create and store your rule documentation. Those systems signal to your agent to load the instructions automatically when a session starts. Rules can also be scoped conditionally now, so they apply only in certain contexts or file types rather than universally.
Skills
A skill is a reusable set of instructions for a specific, repeatable task. Where rules govern general behavior, skills provide your agent with the knowledge to execute specific tasks. The information a skill provides your agent might include example outputs, reference documents, and even a trigger phrase or “code word” that tells the agent when to use or apply it.
Some examples:
“When a sprint ends, summarize velocity, blockers, and carry-over items in this template.”
“When asked for a code review, check for these specific things in this order.”
“When editing copy, flag passive voice, technical jargon, and sentences over 25 words.”
“When creating Figma variants, scaffold all state combinations from the base component using these guidelines.”
Skills are loaded to the agent’s context window when they’re relevant, not all the time. That keeps your agent focused. Like rules, the goal is consistency without constant hand-holding. Most agents, like Mistral’s Vibe, Codex, and Claude Code, all support skill files to extend an agent’s capabilities. And because skills are standardized, they are portable, meaning you can provide a single skill to multiple agents across AI platforms.
Comparing Rules and Skills (The Short Version)
| Rules | Skills | |
|---|---|---|
| Purpose | Define how the agent behaves | Define how the agent executes a task |
| Scope | Always active (or conditionally active) | Activated when relevant |
| Examples | Tone, conventions, safety guardrails | Accessibility patterns, code review steps, doc generation |
Tools We’re Looking At
Several tools have built-in ways to create and manage rules and skills:
Claude supports skills natively through its platform, and rules can be set in project instructions or system prompts. Users can use the chat interface to create a skill, without needing to know the syntax.
OpenAI Codex supports custom instructions that shape agent behavior across sessions. Similarly to Claude, you can create, edit, and refine skills right in a chat window to remove the entry barrier to skill-writing.
Cursor has a dedicated rules system with support for global and conditional rules.
Windsurf offers a similar rule structure with fine-grained scoping options.
The specific tool matters less than the concept. If your agent platform supports persistent instructions, you can apply these ideas regardless of which one you use.
Next Steps in Our Research
At Sparkbox, we’ve begun testing rules and skills across disciplines on our team, but there’s a lot left to explore. Some of the questions we’re still working through:
Breadth Across Disciplines
What rules and skills are most valuable outside of development? What does a useful skill look like for a designer, a PM, or a content strategist? Are there patterns that translate across roles, or does each discipline need its own approach?
Curation and Quality
How do we know when a rule or skill is actually working? What makes a skill well-written versus one that the agent misapplies or ignores? How do we test and validate them before sharing across a team?
Shared Libraries
What would a vetted, cross-discipline library of rules and skills look like for a team? How do we maintain it as tools and workflows evolve?
These are the questions guiding our next phase of exploration. Our goal is to move from individual experimentation toward a shared, curated resource the whole team can use and build on.
What Are You Curious About?
Ask Sparkbox about how AI might be valuable to your organization.
Watch this space for monthly updates on what we’re exploring, what we’ve learned, and what’s next. And if you have ideas or questions, we’d love to hear from you.
