fugu-plan

星标0

分支0

更新时间2026年6月22日 13:42

Fugu-lite Thinker planning for a coding task before any edits. Classifies the task, inspects only relevant files, and produces an execution plan (objective, files, constraints, steps, test strategy, risks, rollback, verification commands, and a ready-to-use worker prompt). Does NOT edit files. Use before starting a non-trivial change.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

来源

jeongsk

jeongsk/fusion-fugu-coding

打开 GitHub 仓库查看创作者相关仓库

下载

Fugu-lite Plan (Thinker)

Plan a coding task before implementation. This is the Thinker role: think, don't build. Do not edit any files in this skill.

Input

$ARGUMENTS is the task to plan (e.g. "Add refresh-token retry for expired sessions"). If empty, ask for the task in one short prompt.

Process

You may run this inline or delegate to the fugu-thinker subagent via the Task tool (preferred for an isolated, focused plan). Either way:

Understand the task. If genuinely ambiguous in a way that changes the plan, state your assumption explicitly and proceed — don't stall on questions.
Inspect relevant files only (grep/glob/read). Never read secret files.
Classify the task: bugfix | feature | refactor | test | docs | architecture.
Detect the project (package manager from lockfile, test/lint/typecheck/build scripts) so verification commands are real. You may run ${CLAUDE_PLUGIN_ROOT}/scripts/detect-project.sh.
Produce the plan using the template at ${CLAUDE_PLUGIN_ROOT}/skills/fugu-plan/templates/execution-plan.md.

Principles

Smallest change that fully solves the task. Keep public APIs stable unless the task requires otherwise.
Make risks and a rollback path explicit.
Verification commands must be concrete and runnable.
Do not edit files. The deliverable is the plan.

Output

Emit the filled-in execution plan. End with a Suggested Worker Prompt: a ready-to-paste instruction for Claude Code (or /fugu-fix) to execute the plan.

同仓库更多 Skills

同仓库

fugu-fix

jeongsk/fusion-fugu-coding

Controlled Fugu-lite coding workflow for bug fixes and feature work — plan (Thinker) then implement (Worker) then verify (Verifier) then Fusion review, with a bounded repair loop. Makes minimal edits; never commits, pushes, or touches secret files. Use to actually carry out a change end to end with guardrails.

2026-06-220

fugu-verify

jeongsk/fusion-fugu-coding

Fugu-lite Verifier for the current changes — inspects git status/diff, detects the package manager and scripts, runs typecheck/lint/test/build where available, scans for secrets, runs a Fusion-lite review, and produces an approve / request_changes / needs_human_review decision. Does NOT implement new changes unless the user explicitly asks. Use to check work before committing.

2026-06-220

fusion-review

jeongsk/fusion-fugu-coding

Fusion-lite multi-agent code review of the current git diff or a provided patch/file. Spawns bug, test, security, and architecture reviewers, then a judge merges findings into one structured decision. Use for PR review, pre-merge checks, final verification, and bug/security/test sweeps.

2026-06-220

name	fugu-plan
description	Fugu-lite Thinker planning for a coding task before any edits. Classifies the task, inspects only relevant files, and produces an execution plan (objective, files, constraints, steps, test strategy, risks, rollback, verification commands, and a ready-to-use worker prompt). Does NOT edit files. Use before starting a non-trivial change.
disable-model-invocation	true
allowed-tools	Read, Grep, Glob, Task, Bash(git status:), Bash(git diff:), Bash(git log:), Bash(find:), Bash(ls:*)

Fugu-lite Plan (Thinker)

Plan a coding task before implementation. This is the Thinker role: think, don't build. Do not edit any files in this skill.

Input

$ARGUMENTS is the task to plan (e.g. "Add refresh-token retry for expired sessions"). If empty, ask for the task in one short prompt.

Process

You may run this inline or delegate to the fugu-thinker subagent via the Task tool (preferred for an isolated, focused plan). Either way:

Understand the task. If genuinely ambiguous in a way that changes the plan, state your assumption explicitly and proceed — don't stall on questions.
Inspect relevant files only (grep/glob/read). Never read secret files.
Classify the task: bugfix | feature | refactor | test | docs | architecture.
Detect the project (package manager from lockfile, test/lint/typecheck/build scripts) so verification commands are real. You may run ${CLAUDE_PLUGIN_ROOT}/scripts/detect-project.sh.
Produce the plan using the template at ${CLAUDE_PLUGIN_ROOT}/skills/fugu-plan/templates/execution-plan.md.

Principles

Smallest change that fully solves the task. Keep public APIs stable unless the task requires otherwise.
Make risks and a rollback path explicit.
Verification commands must be concrete and runnable.
Do not edit files. The deliverable is the plan.

Output

Emit the filled-in execution plan. End with a Suggested Worker Prompt: a ready-to-paste instruction for Claude Code (or /fugu-fix) to execute the plan.