Claude Code Guide
Eight battle-tested moves that combine gstack + ECC + Auto Agent.
Eight battle-tested moves built around three companion tools — gstack commands, ECC Skills, and Auto Agent. Each one is a concrete scenario, the right way to combine commands, and why that combination wins.
The very first message of every new conversation should be this one. Without it, ECC Skills won't auto-detect, and skills like /tdd-workflow, /systematic-debugging, or /verification-loop will only fire when you type them in by hand — twice the work for the same result. Once /using-superpowers has run, Claude reads keywords like "write tests", "fix this bug", or "run verification" and pulls in the matching skill automatically.
Each tool brings its own angle, so run them all at once instead of in sequence. In a single message, fire /review (gstack engineer scoring) + /security-review (ECC deep security) + code-reviewer (Auto quality check). They're independent, so the total wait equals the slowest one. Running them serially is just wasted minutes.
Once Claude finishes writing code, Auto Agent has already run code-reviewer / security-reviewer / build-error-resolver for you. Don't type /review on top — that's tokens for no extra info. Save the manual deep commands for milestones (right before merging a PR): /cso (OWASP Top 10 + STRIDE threat modeling — Auto Agent does not run this for you) and /qa (real browser-click testing).
/office-hours feeds /autoplan. Vague input → vague plan. Use this one-line frame: "I want to build [X] for [user Y] to solve [pain point Z]; the known constraints are [A, B]." Miss one of the three and you'll get questioned back.
Before signing off, write your current progress into PROJECT_STATE.md (gstack ships a template). Next morning, the first message of the new conversation is that file plus /using-superpowers, and Claude is right back where you left off. After every milestone, run /retro and write the sticking points into RETRO.md — read it before starting the next feature so you don't trip on the same things twice.
Three default tiers: subagents (the auto-fired code-reviewer / security-reviewer / etc.) run on Haiku 4.5 — saves 70% on cost. Main conversation runs on Sonnet 4.6. Only switch to Opus for deep-reasoning moments like /cso or architect, then switch back. Inside /autoplan, the CEO / engineering / design viewpoints already pick the right model for each role — you don't need to manage that.
Context windows have limits. The 9-stage flow has three natural break points: after stage 2 planning (PRD is on disk), after stage 4 development (code is committed), and after stage 6 review (report is filed). Before switching, make sure all the key state lives in files — the new conversation picks up via PROJECT_STATE.md plus your committed code. Claude auto-compacts when context hits 50%, but switching deliberately is cleaner.
Jumping straight to /ui-ux-pro-max is a common mistake — you get one design with nothing to compare against, which isn't really a decision. Right order: /design-consultation (clarify product positioning) → /design-shotgun (diverge into 3–5 options at once) → pick a direction → /ui-ux-pro-max (curate palette and fonts from the 67-style library) → /design-html (HTML preview) → /frontend-design (production code).
Vexilo is the missing index for Claude Code — 31 agents, 125 commands, 123 skills and more, organized by scenario. Browse free; unlock search, filters and one-click CLAUDE.md export with a one-time $19 licence.