project-compass - SKILL.md Agent Skill

name: project-compass description: Routes any project to the right next workflow: clarification, documented alignment, TDD implementation, PRD synthesis, debugging, prototyping, code review, deep reading, or Socratic discussion. Never limits routing to student/coursework contexts.

Project Compass

Purpose

Act as a thin router. Do not replace the downstream skills — hand off as soon as the route is clear.

Identify the user's current stage, then route them to the best next skill. Works for any project: personal, course, work, open-source, professional, production, research, or exploratory.

Do not use the user's identity, role, seniority, or learning status as an applicability filter. If memory says the user is a student, use that only to tune explanation style, not to classify the project as coursework or reject work/professional contexts.

Keep responses brief and practical. Prefer Chinese when the user is speaking Chinese.

Workflow

1. Identify the stage

Ask at most three short questions if the route is not obvious:

What do you need right now: clarify, implement/fix, debug, review, produce a document, or read/discuss/learn?
Is there already a repo or codebase? (skip if the request is reading/learning-related)
Is this a one-shot task or an ongoing project?

If the route is already obvious from the user's message, skip questions. Do not ask whether the project is a student project unless the user specifically mentions coursework, class deliverables, teachers, grades, or campus constraints.

2. Route

Use this decision table:

Clarify & plan:

Vague idea, unclear scope, unsure what to build → grill-me
Ongoing repo, repeated confusion about terms, modules, or decisions → grill-with-docs

Build & fix:

Clear feature or bug, wants reliable implementation or test-first flow → tdd
Working idea but wants a quick throwaway implementation to validate → prototype
Has a diff/PR and wants quality review → code-review or review

Diagnose & verify:

Bug is hard to reproduce, performance regression, or unexpected behavior → diagnose
Wants to confirm a change works by running or observing the app → verify

Document & organize:

Needs a requirements doc, project brief, report seed, or structured summary → to-prd
Has a PRD or design and wants structured dev tasks → to-issues
Fixed a complex bug and wants a structured postmortem / 复盘笔记 → postmortem-note

Quality & meta:

Wants to audit, review, or improve a Claude Code skill → skill-review
Wants to prioritize / classify issues, bugs, tasks, or risks → triage

Infra & automation:

Wants long-running autonomous task execution → autonomous-agent
New or existing repo needs a CLAUDE.md scaffold → init
Needs a repeatable SOP for a multi-step workflow → sop-creator

Refactor & improve:

Has working code but architecture or design quality needs improvement → improve-codebase-architecture

Security:

Wants a security-focused review of code or design → security-review

Read & internalize:

Has material (book/article/video/course) and wants structured deep reading → actor-reader
Has consumed content and wants to discuss, challenge, teach, or connect it → socratic-discuss

3. Semi-automatic behavior

If the route is clear and low-risk, explain the reason in one sentence and invoke the target skill.

If the route may create persistent files, tracker items, repo changes, or external issue/PR content, pause briefly and explain what will happen before invoking.

If two routes are both plausible, present the top two choices with a one-line tradeoff and let the user choose.

4. Target skill not found

If the target skill is not installed, say so briefly and offer alternatives:

"[skill-name] 未安装。要我换一种方式处理，还是先安装它？"

Do not fall through to general conversation silently.

5. No-match exit

If the user's intent does not fit any route above, say so briefly and continue the conversation normally. Do not force-fit.

Examples of non-matching requests: tiny one-line edits, asking about tooling unrelated to a project, casual conversation not requiring structured workflow.

Examples

"I want to build a lost-and-found app but the features are messy." → grill-me
"Our project already has code, but user/order/item terms keep drifting." → grill-with-docs
"Help me implement login, test first." → tdd
"I have an idea for a CLI tool, can you quickly hack it together?" → prototype
"The API keeps returning 500 after that deploy, I can't figure out why." → diagnose
"Can you check my local changes before I commit?" → code-review
"Can you review PR #123 before I merge?" → review
"The teacher wants a project requirement summary tomorrow." → to-prd
"This production bug keeps coming back after deploys." → diagnose
"Review this auth module for vulnerabilities." → security-review
"I want to deeply read and internalize《政治学通识》." → actor-reader
"我刚看完 ACTOR 框架的视频，帮我讨论和理解它。" → socratic-discuss
"帮我用 ACTOR 框架读这篇论文。" → actor-reader
"我对这个观点不太确定，帮我挑战一下。" → socratic-discuss
"帮我把这次修复沉淀为复盘笔记。" → postmortem-note
"审查一下这个 skill 的质量。" → skill-review
"帮我分类这些 bug 的优先级。" → triage
"帮我写个多步骤的自动化 SOP。" → sop-creator
"这段代码能跑但架构太烂了，帮我重构。" → improve-codebase-architecture

Guardrails

Do not answer with a long methodology lecture when a clear route exists.
Do not force setup or scaffolding on tiny or one-shot tasks.
Do not rewrite the downstream skills; hand off as soon as the route is clear.
Do not assume the project is student/coursework — treat all projects equally.
Never reject a route because the project is not a student project.
Never infer project type from the user's saved profile; infer it only from the current request and visible project context.