GigiKit Guides

Quick Start: Your First Project

This guide walks through a real end-to-end example: adding a user authentication feature using GigiKit’s /gk:cook skill. You’ll see how CLAUDE.md bootstraps the session, how the orchestrator delegates to sub-agents, and how the plan-implement-test-review pipeline runs automatically.

Step 1: Open Claude Code

Navigate to your project root and launch Claude Code:

claude

Claude immediately reads CLAUDE.md and loads the orchestrator role. It knows to delegate planning, testing, and review to specialized sub-agents rather than doing everything inline.

Step 2: Invoke the Cook Skill

The /gk:cook skill is GigiKit’s primary entry point for feature work. Give it a natural language task:

/gk:cook "Add JWT-based user authentication with login and logout endpoints"

Cook detects the intent, selects interactive mode by default, and begins the workflow pipeline:

[Intent Detection] → [Research] → [Review] → [Plan] → [Review] → [Implement] → [Review] → [Test] → [Finalize]

Step 3: Review the Research Findings

Cook spawns a researcher sub-agent to gather context about your codebase and the task. It pauses at the first review gate and presents findings:

✓ Step 1: Research complete - 3 relevant patterns found
  → Existing auth middleware in src/middleware/
  → Database schema supports users table
  → JWT library already in package.json

Proceed to planning? [Y/n]

Type Y to advance. This human-in-the-loop gate ensures you catch wrong assumptions early.

Step 4: Approve the Plan

The planner sub-agent generates a structured plan saved to ./plans/:

plans/260305-0912-jwt-authentication/
├── plan.md
├── phase-01-database-models.md
├── phase-02-auth-endpoints.md
├── phase-03-middleware.md
└── phase-04-tests.md

Cook shows a summary and pauses again:

✓ Step 2: Plan created - 4 phases, ~2h estimated
  → phase-01: User model + migration
  → phase-02: /login, /logout, /refresh endpoints
  → phase-03: JWT verification middleware
  → phase-04: Integration tests

Approve plan and begin implementation? [Y/n]

Step 5: Watch Implementation Run

After approval, a fullstack-developer sub-agent executes each phase sequentially. You see progress as files are created:

✓ Step 3: Implementation complete
  → src/models/user.ts (created)
  → src/routes/auth.ts (created)
  → src/middleware/jwt-verify.ts (created)
  → 847 lines added across 3 files

Step 6: Tests and Review Run Automatically

Cook then delegates to tester and code-reviewer sub-agents — you do not need to invoke these manually:

✓ Step 4: Tests passed - 24/24, coverage 91%
✓ Step 5: Code review complete - score 9.2/10, 0 critical issues

Step 7: Finalize

Cook’s mandatory finalization step syncs the plan, updates ./docs/, and offers to commit:

✓ Step 6: Finalized
  → plan.md updated with completion status
  → docs/codebase-summary.md updated
  → Ready to commit? [Y/n]

What Just Happened

The full delegation chain looked like this:

Cook Workflow Delegation 
graph TD
A[“/gk:cook task”] —> B[Intent Detection]
B —> C[researcher sub-agent]
C —> D[“Review Gate”]
D —> E[planner sub-agent]
E —> F[“Review Gate”]
F —> G[fullstack-developer sub-agent]
G —> H[tester sub-agent]
H —> I[code-reviewer sub-agent]
I —> J[project-manager + docs-manager]
J —> K[git-manager sub-agent]

Each agent owns a specific responsibility and hands off cleanly. This is the core GigiKit pattern: the orchestrator coordinates, specialists execute.

Next Steps