Skills in a harness
Skills are reusable capability packages that combine instructions, tools, constraints, and completion criteria for a specific type of job. They make task execution more reliable than one universal prompt.
A skill is a pre-defined way of doing work inside the harness. It usually includes a domain-specific objective, curated toolset, step order, guardrails, and output format. In practice, skills function like operational playbooks the model can execute repeatedly with less drift.
Skills improve completion by reducing ambiguity at the moment of action. Instead of letting the model invent a workflow each time, the harness routes the request to a tested pattern. This improves speed, consistency, and safety while lowering token waste from unnecessary exploration.
- Intent framing: what success looks like for this task category.
- Allowed actions: which tools are available and which are blocked.
- Execution policy: order of operations, retries, and stop conditions.
- Output contract: required structure, evidence, and verification checks.
- Higher reliability: fewer hallucinated steps and fewer malformed outputs.
- Better safety: high-risk actions stay gated inside approved workflows.
- Faster delivery: specialized context shortens decision time per loop.
- Easier evaluation: repeated task types can be measured against stable criteria.
At intake, the harness classifies the request and selects a skill. During planning, the chosen skill narrows context and tool choices. During execution, the harness enforces that skill’s policies and sequencing. During reflection, results are checked against the skill’s completion rubric. If requirements are unmet, the skill defines recovery behavior or escalation paths.
Common skill groups include coding change execution, research synthesis with citations, incident triage, workflow automation, and policy-aware content operations. Each family can expose different tool budgets, approval requirements, and validation steps while still using the same base harness runtime.
Keep each skill narrow enough to be predictable, but broad enough to handle normal edge cases. Define explicit inputs and outputs, include robust failure handling, and make verification mandatory before completion. Treat skills as versioned assets: test them, measure real task outcomes, and retire ones that underperform.