A prompt is a single shot. Every time you ask Claude to do something complex you re-explain the rules, drag in the same context, and pay tokens for repetition. Agent Skills flip the model: a skill is a folder containing metadata, scripts, and examples. Claude lazy-loads only the metadata first — deciding whether to invoke deeper assets later. The result is reusable expertise that survives across sessions, instead of one-shot text that dies the moment the window closes.

Every disciplined project starts with structure. ck init scaffolds the .claude/ directory, drops in agent.md and claude.md, and seeds a rules/ folder. Those files are not decoration — they are the harness. Agent.md describes who Claude is on this project, claude.md captures conventions, and rules become the project-specific guardrails Claude must respect on every call. Treat ck init as the moment you stop chatting and start engineering.

ck init

The most powerful skill is the one that builds other skills. skill-creator uses Claude itself to scaffold a new skill folder — SKILL.md, scripts, references — from a short description. The pattern is recursive in the best way: when you spot a recurring fix, promote it into a rule. When the rule grows hands and feet, promote it into a skill. Meta-tooling collapses the distance between "I keep doing this manually" and "Claude does it for me reliably."

claude → /skill-creator

Context is finite. The harness assumes a 40% baseline of always-loaded scaffolding plus a working window that degrades as it fills. At roughly 70% utilization Claude auto-compacts, but auto is too late — you have already lost fidelity in the middle. Make compaction proactive: /compact after every feature, /clear when starting a fresh thread, and offload long discussions into BD.md so the session can rebuild context on demand instead of carrying it forever.

Compact after every sprint, every feature — do not wait for auto.

/compact
/clear

ck brainstorm turns a vague idea into a Business Definition file through interactive Q&A. Claude probes for scope, success criteria, edge cases, and constraints — and the answers crystallize into BD.md. The danger is reflexively accepting whatever option Claude marks as "recommended." A click takes a second, but the disaster trailing behind a wrong default takes weeks to undo. Read every option, push back when something feels off, and treat the BD as a contract before any code is written.

A click takes a second; the disaster trailing behind takes weeks to fix.

ck brainstorm <topic>

ck plan reads BD.md and produces a phased implementation plan with explicit todos. The plan is structured so a token-exhausted session, or even a fresh Claude tomorrow, can pick up exactly where it stopped. But never trust the plan blindly. ck red-team spins adversarial agents at it — hunting for assumption failures, hidden risks, and missing edge cases. The findings get accepted, rejected, or rewritten back into the plan. By the time you cook, the plan has already survived its first attack.

Never trust the plan blindly — always red-team it.

ck plan BD.md

ck red-team plan.md

Reality is multi-IDE. Claude Code, OpenCode, and Antigravity all ship their own built-in skills, and when you let the agent auto-pick from a crowded room you get reward-randomness — whichever skill happened to score highest, not the one you actually want. The fix is mechanical: invoke skills explicitly with /<skill-name>. Force the choice. Auto-pick optimizes for the agent's confidence, while explicit invocation optimizes for your intent. Disciplined engineers always force.

Best practice — force the skill, do not let it auto-pick.

claude → /<skill-name>

The .claude/rules/ folder is where corrections become permanent. The first time Claude breaks a convention you write a rule. The second time, the rule fires before the mistake. Rules are short, declarative markdown files — one concern per file, kebab-case names, easy to grep. When a rule starts feeling like a procedure, promote it into a skill. Rules are the cheapest, most underrated layer of the harness; they catch the boring failures that would otherwise eat your week.

.claude/rules/*.md