Two Claude Code skills that interrogate you until your task is shaped into a precise, executable spec — a depth
/goalprompt or a breadth Dynamic Workflow. A design coach, not the engine.
🇹🇼 繁體中文讀者 → 看白話中文說明：README-tw.md（這是什麼、裝了有什麼好處、怎麼用，全中文）
This is a designer, not a feature. It does not replace — and is not required to use — Claude Code's /goal or its Dynamic Workflows. These skills are front-end coaches: they interrogate you and emit the precise prompt/spec you then feed to those features. The hard part was never running /goal or a workflow — it's writing the spec that goes into them. That's what this designs.
它是設計器，不是功能本身。不取代、也不需要它才能用 /goal 或 Dynamic Workflows。這兩支 skill 是前端引導器：反問你、產出精準的 prompt/規格，再由你貼進那些功能。難的從來不是「跑」/goal 或 workflow，而是「寫出餵進去的規格」——這就是它設計的東西。
A vague task can be shaped two ways. This repo ships one skill for each axis:
| Skill | Axis | Shapes your task into… | Reach for it when |
|---|---|---|---|
goal |
Depth (深度) | a precise /goal prompt — one agent + a rubric, iterating to a quality bar |
one thing to get right, deep, and converged |
workflow-shaper |
Breadth (廣度) | a runnable Dynamic Workflow — many fresh-context agents in parallel, cross-checked | the same check/change across many independent units |
handoff |
Continuity (接續) | a self-contained startup prompt for a fresh session — plus an optional one-click "open a new Remote Control session and send it" | a long task has to move to a new session (e.g. context is filling up, or you're away and want to watch it from the phone) |
One-second test for which to use: count how many units. One thing to perfect → /goal. The same check across many things → workflow-shaper. Many different things with a dependency order → that's spec decomposition, see the companion repo spec-sonar below. They compose — a workflow is the for-loop, a goal is the loop body.
判斷用哪個的一秒法則：數有幾件。一件做到完美 → /goal；同一個檢查套到很多件 → workflow-shaper；很多件「異質且有依賴順序」→ 那是規格分解，見下方姐妹 repo spec-sonar。兩者可組合（workflow 是 for-loop、goal 是 loop body）。
In 2026, Claude Code, OpenAI Codex, and Manus Agent shipped a /goal feature within weeks of each other; Claude Code later added Dynamic Workflows (background multi-agent orchestration). Both loop or fan out agents toward a target. Anthropic's research showed that without a precise target, LLMs hit context anxiety when their context window fills — they shortcut to a wrap-up and stop, even mid-task.
But these features only work when the spec is precise. Given a vague prompt, the AI invents its own definition of "done" and quits early — or, for a workflow, fans out on the wrong shape and burns tokens. Given a precise one, it iterates for hours / parallelizes cleanly.
The hard part isn't using /goal or a workflow. It's writing the spec that goes into them. These skills automate that, and refuse to let you off the hook with vague answers.
Paste this into your Claude Code session and the agent installs both skills into the right place:
Read https://raw.githubusercontent.com/dragon375014/goal-workflow-designer/main/INSTALL.md
Then install both skills (goal + workflow-shaper) into my user-level ~/.claude/skills/ by following it.
Show me a diff/summary before writing any file, and don't overwrite a SKILL.md I've edited without asking.
The agent fetches INSTALL.md, places skills/goal/ and skills/workflow-shaper/ under ~/.claude/skills/, and appends the two triggers to your ~/.claude/CLAUDE.md.
git clone https://github.com/dragon375014/goal-workflow-designer.git /tmp/gwd
cp -r /tmp/gwd/skills/goal ~/.claude/skills/goal
cp -r /tmp/gwd/skills/workflow-shaper ~/.claude/skills/workflow-shaper
cp -r /tmp/gwd/skills/handoff ~/.claude/skills/handoffThen add the two trigger blocks (see INSTALL.md Step 3) to ~/.claude/CLAUDE.md. These are user-level skills — they work across all your Claude Code projects. Output goals are saved under ~/.claude/goals/<name>/.
⏫ Upgrading from
skill-to-goal(the old single-skill repo name)? The repo was renamed and restructured: the goal skill moved from the repo root toskills/goal/, andworkflow-shaperwas added. Re-run Option A, orgit pulland re-copy.
/goal <one-line task description> # full Q&A from scratch
/goal --use <name> # load a saved goal
/goal --resume <draft-name> # resume an interrupted Q&A
It walks you through five elements (outcome, verification, constraint, iteration policy, error handling), plus a six-step rubric SOP for subjective work, and outputs a goal prompt that keeps the agent iterating for hours instead of quitting in three minutes. Natural language triggers too (幫我寫個 goal prompt).
⚠️ /goalwith no arguments is intentionally unsupported — vague triggers produce vague Q&A.
Describe a many-unit task, or say "use a workflow." It runs a WORTH-IT gate (5 criteria → green/yellow/red) before anything else:
- 🟢 green → recommends a fan-out shape (parallel / pipeline / map-reduce) and emits a paste-ready
workflow ...prompt + a draft meta block you can save as a/<name>command. - 🟡 yellow → offers a middle path (split into short workflows / inline-run one unit first).
- 🔴 red (single unit, needs intermediate context, needs mid-run sign-off) → refuses to emit a workflow prompt and routes you back to a plain skill or
/goal.
It never runs the workflow — it shapes the spec and hands it back. See skills/workflow-shaper/SKILL.md.
A handful of Claude Code skills do prompt design (prompt-architect, prompt-improver, prompt-master). They are general-purpose. This one is laser-focused on the /goal execution model:
| Feature | this repo | Generic prompt skills |
|---|---|---|
Targets /goal's five-element framework |
✅ | ❌ (general frameworks like CO-STAR, RISEN) |
| Builds a rubric for subjective tasks | ✅ (6-step SOP) | ❌ |
| Context window health gate | ✅ (Phase 0a) | ❌ |
| Goal-eligibility check (3-light) | ✅ (Phase 0c) | ❌ |
| KICK-OFF self-cue (agent actually starts, no "continue" needed) | ✅ | ❌ |
| Handoff packaging for context-strained sessions | ✅ (--resume) |
❌ |
| Codebase scan via isolated subagent | ✅ | ❌ |
| A breadth counterpart for many-unit tasks | ✅ (workflow-shaper) |
❌ |
| Bilingual term parentheticals (EN + 中文) | ✅ | varies |
These are codified in the skills and govern every Q&A round:
- Interrogate vague answers. "better / smoother / more professional" → the skill must follow up for a measurable standard. Accepting vague answers is a failure.
- Smart skipping. Scan everything already collected; skip known elements.
- 1–3 questions per round. Long lists make users abandon.
- Cite examples to seed thinking — but tell users not to copy verbatim.
- Subjective tasks require a rubric. Without one, the reviewer agent has nothing to judge against.
- Context window health above all (Phase 0a gatekeeper).
- Not every task is
/goal-shaped (Phase 0c) — and not every task is worth a workflow (workflow-shaper's WORTH-IT gate). Both refuse rather than ship a doomed run. - KICK-OFF is a self-cue. After printing the kick-off block, the skill makes a tool call in the same turn — text alone doesn't reliably start the agent loop.
- Bilingual terminology. First mention of any term writes
English (中文).
A novel mechanism — see docs/handoff-pattern.md. When a session is already context-strained, finishing the Q&A then running /goal compounds the problem. The skill packages in-progress state into handoff.md and tells the user to /clear + /goal --resume <draft-name> so the actual run lands on a full-context window. Generalizes to any long-running skill.
This repo is the shaping layer of a five-public-repo AI-dev toolchain by dragon375014. Full topology, routing rules, and canonical skill ownership → ECOSYSTEM.md
One-command install (drops all five tools into the right place):
npx specmit initSiblings: spec-sonar · specmit · claude-skills-governance-meta · agent-work-board — each stands alone; install only what you need.
This repo is the execution side (how to run one task deep, or many units wide). spec-sonar is the spec side (what to build): idea-to-spec converges a vague product idea into a spec via dark-zone Q&A, and goal-decomposer compiles that spec into a dependency-ordered goal graph — whose G*.md files use the same five-element format this repo's /goal produces, so they are directly executable via /goal-style KICK-OFF.
| You want | Use | Repo |
|---|---|---|
| Converge a full product idea into a spec | idea-to-spec |
spec-sonar |
| Split a finished spec into dependency-ordered units | goal-decomposer |
spec-sonar |
| Polish ONE task into an iterable goal prompt | /goal |
this repo |
| Run the same check across N independent units | workflow-shaper |
this repo |
Full chain: idea-to-spec → goal-decomposer → goals/G*.md → execute each via /goal-style KICK-OFF (depth), fan out homogeneous sub-tasks via workflow-shaper (breadth).
This repo shapes what to do (the task spec). Its governance companion shapes how it's built (the structure):
Defensive + offensive governance patterns: an architecture-completeness gate (catches "I want to build X" before it's built), cross-layer trace-lock (lock a write→display chain so "changed A, forgot B" can't happen), and runnable defensive linters.
They combine cleanly:
workflow-shaperhere can be dispatched by that repo'sarchitecture-completeness-guardian(the L0 gate decides "this is a many-unit job" → routes to the breadth shaper).- A
/goalor workflow you design here is executed under that repo's trace-lock + governance guard — so a multi-unit change stays SSOT-locked instead of drifting. - Mental model: this repo = the brief; that repo = the guardrails. Use one for sharper task specs; add the other when your project is big enough to need structural governance (it ships a fitness check that tells you when).
If you only installed one, that's fine — each stands alone. But for a solo developer running long AI sessions on a real codebase, the pair covers both "shape the work" and "don't let the work corrupt the structure."
The methodology draws from Anthropic's research on context anxiety and the implementer/reviewer architecture, Anthropic's frontend-design skill case study (the museum experiment → the 6-step rubric SOP), Karpathy on evaluation > prompt engineering, and the community insight that /goal succeeds or fails entirely on spec precision. The breadth half draws from Claude Code's Dynamic Workflows (background multi-agent orchestration). A transcript is in examples/source-transcript.md.
PRs and issues welcome. A few principles:
- Don't add elements to the five-element framework. It's stable on purpose. Open an issue first.
- Don't widen
workflow-shaper's WORTH-IT gate to auto-approve. The point is that it refuses cheap tasks. - Add example outputs to
examples/. - Translations welcome (currently EN/繁中).
- Test before PR — run the skill end-to-end on your machine.
MIT. Do whatever you want with it; attribution appreciated but not required.
- Claude Code skills documentation
- Claude Code Dynamic Workflows
- awesome-claude-skills
- prompt-architect — general prompt structurer
I'm a solo dev running a coffee brand who taught myself to build this toolchain with AI. I write up the reasoning (and the scars) at coffeeshooters.com/content — start with Magic Defeats Magic: how the /goal skill interrogates a vague task into a precise goal prompt. Two-week newsletter, no spam — subscribe on any article.
If either skill saved you a debugging session or a wasted /goal run, a ko-fi is appreciated but not expected. Everything here is MIT and free to use unconditionally.
