| name | sequential-thinking |
| description | Break a multi-step task into numbered reasoning steps before acting. Used by taw-kit planner agent to decompose ambiguous Vietnamese prompts into executable phases. Vietnamese trigger phrases: "suy nghĩ từng bước", "phân tích kĩ", "chia nhỏ ra".
|
sequential-thinking
Forces deliberate, numbered reasoning when a problem is too fuzzy to one-shot. Prevents plan holes.
When to invoke
- Intent classification confidence is low (user text is ambiguous)
- Requirements conflict (user asks for auth + no login in the same sentence)
- Estimating effort for a feature the orchestrator has not built before
- Whenever the output would otherwise be "let me just try and see"
Method
Produce a numbered list where each step satisfies:
- One action verb per step. "Check", "Decide", "Write", "Invoke".
- Observable output. What file/state changes after the step?
- Cheap before expensive. Verify assumptions before spawning agents or running npm install.
- Stop condition. Each step has a pass/fail check before the next step is considered.
Rendering contract
Internal reasoning. The user does NOT see these steps unless they hit an error — then surface the last 3 steps as context for $taw fix.
Example
User prompt: "xây blog cá nhân, kiểu như medium, có login, dùng chung tài khoản Google"
Reasoning:
- Check intent:
blog with social login — multi-user or single-user?
- Spot conflict: "cá nhân" (solo) vs "chung tài khoản Google" (multi). Decide: ask user.
- If solo → skip Google OAuth, use Supabase magic-link for author only.
- If multi → add Google provider, but this raises scope ≥ CRM level.
- Default assumption: solo author + Google OAuth for SSO convenience only.
- Load
taw clarify template Q16 (blog storage) + custom Q (Google login for you or for readers?).
Only step 6 is visible — prior steps stay in the reasoning scratchpad.
Anti-patterns
- Don't produce "plan of plan" loops. If the reasoning itself needs sequential-thinking, stop and ask the user.
- Don't fabricate deadlines ("Step 7 takes 5 minutes"). Time estimates belong in the plan bullets, not the reasoning trace.
- Never expose intermediate reasoning to the user unless debugging.
