From v3 onward, /moai’s default routing is Analyze-First. It classifies request meaning rather than matching English keywords. So you get the same routing quality in any language — whether you ask in Korean “로그인 버그 고쳐줘” or in English “fix the login bug”, intent analysis routes both to the same workflow.

This page walks through the five stages Analyze-First uses to turn a request into execution.

Every request — /moai subcommand or natural language, any input language — flows through one ordered pipeline. Each stage takes the previous stage’s result as input.

flowchart TD
    A["① Intent analysis
Classify intent language-independently"] --> B["② Context-sufficiency check
Insufficient → Socratic interview"]
    B --> C["③ Execution plan composition
Skill·agent·workflow chain"]
    C --> D["④ Approval gates
Implementation kickoff approval, etc. human gates"]
    D --> E["⑤ Execute → verify → iterate
Until acceptance criteria met"]

Classify the request’s intent independently of language. Whether input is Korean, English, or Japanese, technical signals serve only as context for stage ③ and are NOT erected at routing forks. “이거 고쳐줘” and “fix this” both classify as the same intent (fix).

After intent classification, check whether context is sufficient to execute the request. If insufficient, run Rule 5 Context-First Discovery’s Socratic interview via AskUserQuestion to pose clarification questions. If sufficient, proceed to the next stage.

With context in hand, compose an execution plan. Decide which skills to load, which agents to spawn in what order. For non-trivial work, this plan is surfaced to the user before execution (Approach-First). Here Phase 4 orchestration mode is also chosen, deciding whether parallel fan-out or sequential sub-agents is appropriate.

When the plan is set, pass through the pipeline’s named gates in order. In particular, Implementation Kickoff Approval (plan → run human gate) is NOT an autonomous bypass target — even when plan artifacts are audit-ready, immediately before run-phase entry you MUST obtain the user’s explicit approval via AskUserQuestion. The autonomy axis (semi-autonomous vs. autonomous) chooses what happens AFTER approval passes, not skipping the gate itself.

With approval, run the plan. Verify against acceptance criteria and iterate as needed. When a goal is armed (/moai goal), the goal evaluator decides completion.

The base pipeline passes through four gates in order. Each gate is independent — a FAIL or INCONCLUSIVE stops the chain.

Info

Implementation kickoff approval is score-independent. Even when Plan-audit is PASS-eligible (≥ 0.90), user approval before run-phase entry is NOT skipped. Phase 0.5 SKIP and implementation kickoff approval are different decisions — the former is whether to re-run plan-auditor (automatable), the latter is whether the user enters run (user decision mandatory).

When you enter only natural language without a subcommand, ① intent analysis automatically picks an appropriate workflow.

• Fix intent → fix family workflow
• New feature intent → plan → run → sync pipeline
• Exploration intent → Read-only subagent fan-out

For example, entering just /moai "로그인 버그 고쳐줘" still classifies as “fix” intent and routes to the fix family. /moai "소셜 로그인 추가하고 싶어" classifies as “new feature” and reaches the 3-stage pipeline. Single-word intents like /moai status likewise route appropriately.

• No language friction — No need to memorize English keywords. Request in your native language and get the same routing quality.
• Transparent gates — Before execution, it’s revealed which skills·agents will be called in what order.
• Consistent approval points — Even in autonomous mode, users can block before run entry at least once.

• /moai — The entry command for Analyze-First
• SPEC-based development — The parent frame for the plan → run → sync pipeline
• Harness profiles and evaluation — Gate evaluation and Tier classification