← Back to insights

Skills and Harnesses: Making AI Results Consistent

Why the same request can produce different quality from day to day, and how skills, project context, and a harness make AI output something you can rely on.

In short

  • Inconsistent AI is usually not less capable, just short on context and guardrails.
  • A skill packs a repeatable way of working into instructions that load when needed.
  • A harness is the tests, lint, review, and progress notes around the model. That is where consistency is built.

On Monday morning we asked an agent to build a settings page. The result was tidy: our components, our file naming, the right buttons. On Thursday we asked for almost the same thing. It came back with a different style, different buttons, a different folder structure.

Same model. What differed was us: on Monday we explained at length, on Thursday we were in a hurry. That taught us consistency does not come from a smarter model but from how we prepare the work.

The problem isn’t intelligence, it’s memory

A model carries no memory from one session to the next. Without written context, every session is a first day on the job. So the task shifts: not coaxing the model with a perfect sentence, but preparing the same working environment every time.

Skills: recipes loaded when needed

A skill is a folder of instructions (usually a SKILL.md file), sometimes with scripts and examples, that teaches an agent one kind of work: writing a report, checking accessibility, or building components that follow a design system. According to the Agent Skills documentation, only a short name and description load up front. The full content is read when a task needs it, so you can keep many skills without crowding the agent’s head.

  • One clear task, not “everything about frontend”.
  • A description that says when to use the skill.
  • An example of a correct result.
  • A list of things not to do.

A harness: guardrails around the model

A harness is everything outside the model that shapes how it works: the tools it may use, the context it loads, the tests it must pass, and how it reports progress. Anthropic wrote about this for agents that run for hours: split the work, record progress in files, commit changes, and test features before calling them done (Effective harnesses for long-running agents).

The idea is familiar to anyone who has done a shift handover: write down what is finished first, so the next person doesn’t have to guess.

Skillhow to workContextproject rulesModelwritesGuardrailstests · lintReviewhuman↺ repeat until it passes✓ Consistent result
Skill and context prepare the model, automatic guardrails filter the result, a person judges the rest. If it fails, go again.

Our recipe for consistency

  1. Write project rules in one context file such as AGENTS.md: stack, code style, and the commands that run the tests.
  2. Turn repeated work into skills.
  3. Add automatic guardrails: tests, lint, and type checks run before a person sees the result.
  4. Ask the agent to note what is done and what is not.
  5. Keep human review as the last gate. It stays small and focused because the machine already filtered the trivial.

How it feels once it’s a habit

A rushed Thursday no longer produces different quality, because quality no longer depends on how long we explained things that day. For us that is the biggest change: from “hopefully it’s good this time” to “this is the default”.

We bring the same habits to client work, from our development services to the UI kits in the Store. And if you wonder whether all this means people are no longer needed, read Will AI Replace Engineers and Designers?.

Start a project

We reply within one working day.

What do you need?
Budget