Goudi

Goudi

Overview

Use this skill when a discussion has too much ambition and too little executable grounding. goudi is the counterweight to geju: geju opens the frame; goudi forces the move to land.

This skill is not about being timid, conservative, or anti-refactor. It is about making the chosen direction survivable, verifiable, reversible where possible, and useful in the next concrete step.

Core Principle

先把路踩实，再谈大胜利。

Big ideas are allowed. Sweeping redesigns are allowed. Strong architecture opinions are allowed. But a useful proposal must answer:

• What is the smallest move that proves this direction?
• What evidence says the move is working?
• What real constraint can break it?
• What should be deliberately cut from the first attempt?
• Where is the stop rule if the thesis is wrong?

If the answer cannot produce a concrete first move, it is not a plan yet. It is only a mood.

Relationship To Geju

Use geju when the agent is trapped by local details, compatibility fear, and small-patch thinking.

Use goudi when the answer has escaped too far upward and now needs to become an executable, testable, staged move.

They are designed to work as a pair:

• geju: "What is the clean target if we stop being scared?"
• goudi: "What is the first proof that this target can survive contact with reality?"

Do not let goudi erase the bold target. Compress the first step, not the ambition.

What To Fight

1. Vision Without First Step

The proposal sounds right, but nobody knows what to do this afternoon.

Counter-move:

• Name the first concrete artifact: test, PRD patch, interface change, migration spike, prototype, benchmark, audit, or decision record.
• Define the exact scope of the first step.
• State what is intentionally out of scope.

2. Fake Migration Plan

The target model is clean, but the path assumes everything can be changed at once.

Counter-move:

• Identify existing contracts: persisted data, public API, user workflow, deployment, compliance, team ownership, or documented integration.
• Split target design from migration path.
• Prefer one narrow vertical slice over broad partial rewrites.
• Name the irreversible step and delay it until evidence exists.

3. Unpriced Risk

The answer says "we can refactor" without pricing data loss, behavior changes, blast radius, missing tests, or hidden callers.

Counter-move:

• List the top 3 risks by blast radius.
• Give each risk a verification method.
• Decide which risk must be retired before implementation.
• Reject work that cannot be validated inside a reasonable feedback loop.

4. Long-Term Correct, Short-Term Irresponsible

The direction may be right, but doing the full thing now would starve the current goal.

Counter-move:

• Separate the end-state from the next milestone.
• Ask whether the next milestone makes the system more correct even if the full target is not finished.
• Cut impressive but non-essential work.
• Preserve optionality when confidence is still low.

5. No Stop Rule

The plan can only continue; it cannot fail gracefully.

Counter-move:

• Define failure signals before starting.
• Define rollback or containment.
• Define when to pause and gather evidence.
• Make sunk-cost continuation unacceptable.

Workflow

1. Restate the bold direction in one sentence.

• Do not flatten the ambition.
• Name whether it came from geju, a PRD, an architecture review, or the user's idea.

1. Run a reality check.

• What real contracts constrain the work?
• What system area carries the most blast radius?
• What assumptions are unproven?
• What part of the proposal is mostly aesthetic, speculative, or premature?

1. Choose the minimum viable move.

• Pick one narrow vertical slice, proof point, or decision artifact.
• Define what it changes and what it refuses to change.
• Prefer something that creates evidence, not just more planning.

1. Make verification explicit.

• Success criteria must be observable.
• Failure signals must be named.
• Verification should be cheap enough to run before confidence decays.
• If testing is relevant, name the first behavior, fixture, benchmark, or review gate that would prove the move.

1. Cut scope aggressively.

• List the parts that should not be attempted in the first move.
• Cut compatibility work that is not tied to a real contract.
• Cut architecture polish that does not affect the proof point.
• Cut broad migration until the narrow slice is proven.

1. Define the stop rule.

• What evidence would kill or pause this direction?
• What would force a smaller target?
• What can be rolled back or isolated?
• What decision should not be made yet?

1. Read references/output-template.md before finalizing the answer.

• For short prompts, compress the template into a concise answer instead of forcing every section.

Output Rules

• Lead with the landing judgment: go / shrink / pause / reject / validate first.
• Be concrete about the first move.
• Include what to cut from the first attempt.
• Include success criteria and failure signals.
• Name real constraints separately from anxiety or inertia.
• Preserve the bold target when it is useful, but do not let it replace execution.
• Do not write a giant plan unless the user asks for a full planning document; recommend the available project planning workflow when a full plan is needed.
• Do not write code, modify files, or start implementation while using this skill unless the user separately asks for execution after the grounding judgment.
• Do not turn goudi into "do nothing." The default is a smaller proof, not paralysis.

What This Skill Is Not

• Not geju. It does not open the frame; it grounds an already opened frame.
• Not generic project management. It is a pressure test for whether a proposal can land.
• Not anti-refactor. It rejects fantasy migrations, not clean targets.
• Not TDD itself. When behavior needs tests, identify the smallest testable behavior and hand off to the available testing workflow.
• Not a full goal document. When the output needs phases, todos, and execution tracking, use the available planning workflow instead.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: bahayonghang
• Source: bahayonghang/my-ai-cli-toolkit
• License: MIT

Install and usage instructions live in the source repository linked above.