Glossary

A distinct way a skill can be invoked — a case it handles, so different runs take different paths through the same skill file.

Keeping everything needed at once — a term's definition, rules, caveats — under one heading rather than scattered across the skill file.

The cost a user-invoked skill puts on the human — remembering it exists and when to reach for it. The price of human agency, not a bug.

The condition that tells the agent a step is done. Clarity resists premature completion; demand sets how much legwork it forces. Must be observable — not vague like "thoroughly" or "shared understanding."

The always-on cost a model-invoked skill spends: its description, every turn, whether needed or not. With 30+ skills, this becomes a budget problem.

A reference held in context naming out-of-context material. Its wording decides when — and how reliably — the agent reaches for it.

The pruning heuristic: "If I remove this line, what changes in the agent's behaviour?" If nothing changes, the line is a no-op — delete it.

The same meaning stated in more than one place. Costs maintenance and tokens, and inflates the duplicated idea's apparent importance to the model.

Reference living outside the skill system entirely — a plain file any skill can point at, including two user-invoked skills that can't invoke each other.

How finely you split skills. Split further and you spend either context load (more descriptions) or cognitive load (more for the human to track). There's no free lunch.

A compact concept already in the model's pretraining (e.g. "bisect," "fog of war," "tracer bullet") that anchors behaviour in the fewest tokens. Coined words cost more — they carry no free priors.

The work an agent does within a step — exploring, digging, verifying — rather than offloading it back to the user. More legwork = the agent earns its keep.

The path a skill walks from creation to retirement: Draft → Promote → Rename/Reframe → Merge → Deprecate/Remove. Skills that never move are stagnating.

A skill that keeps its description loaded so the agent can fire it autonomously. Always pays context load; appropriate when the agent should reach for it without being told.

Steering by prohibition — naming what not to do drags it into the frame and makes it more available, not less. Cure: state the positive target instead. Exception: hard guardrails where violation has severe consequences.

An instruction that changes nothing because the model already does it by default — you pay context tokens to say what would've happened anyway. Detected by the deletion test.

Steps that follow the current one. When visible, they pull the agent forward — the stronger the pull, the more likely premature completion of the current step.

Ending a step before it's genuinely done because attention slips ahead. Cured first by sharpening the completion criterion, only then by hiding later steps via progressive disclosure.

Pushing reference out of the skill file's top-level and behind a pointer, licensed by which branches actually need it — keeping the top of the file legible and reducing sprawl.

Material referred to on demand: definitions, examples, parameters. Secondary to steps, or the whole content when there are no steps.

Whether a line still bears on what the skill does — either it never mattered, or it drifted stale as the world it describes changed.

A user-invoked skill whose job is pointing at your other user-invoked skills, so the human has one thing to remember instead of many. Example: ask-matt from the trail-map.

Layers of old content that never get cleared because adding feels safe and removing feels risky. The default fate of any skill without a pruning discipline.

Each meaning lives in exactly one authoritative place, so a behaviour change is a change in one place. The antidote to duplication.

A skill that's simply too long, independent of staleness or repetition. Cured by pushing reference down the hierarchy via progressive disclosure and splitting by branch.

The ordered actions the agent performs — the primary tier of a skill's information hierarchy. When a skill has them, they're what the agent reads first.

A skill with its description stripped — reachable only by a human typing its name. Zero context load, but invisible to other skills. Appropriate for deliberate acts like handoff or teach.