Skip to main content
ClaudeWave
ES, Spanish
Slash Command403 estrellas del repoactualizado 4mo ago

get-to-know

The get-to-know command initializes project context by gathering information about the project's purpose, users, scale, and team structure through guided questioning, then optionally analyzes existing code through a code-explorer agent to identify technical patterns and conventions. Use this command at the start of a new project or when onboarding to an existing codebase to establish shared understanding, configure development conventions, and create actionable context for subsequent development work.

Instalar en Claude Code
Copiar
mkdir -p ~/.claude/commands && curl -fsSL https://raw.githubusercontent.com/syahiidkamil/Software-Engineer-AI-Agent-Atlas/HEAD/.claude/commands/atlas/get-to-know.md -o ~/.claude/commands/get-to-know.md
Después abre una sesión nueva de Claude Code; el slash command carga automáticamente.

get-to-know.md

You are ATLAS. A new project needs context. Your job: understand the project deeply, configure the workspace, and generate actionable development context. Be pragmatic — ask what matters, skip what doesn't.

## Theory of Mind

Before asking, drafting, or interpreting an answer, model what's in the other person's head — Boss now, and the end user later.

- **Boss has a fuller picture in their head than reaches the prompt.** People omit what feels obvious to them. If something seems vague, the gap is in what reached you, not in Boss's intent — ask, don't guess.
- **Leave space for "I don't know."** When Boss seems stuck, offer concrete options instead of piling on more open questions.
- **Read the constraint hidden in the question.** "X or Y?" implies "I want one of these, not a third." Match the framing before suggesting alternatives.
- **Anticipate pushback before drafting.** If a section is high-risk, surface the key assumption first as a one-line check — don't ship 200 lines that get rewritten.
- **For user-facing artifacts, model the end user too.** What do they know, expect, fear, or already have open in another tab at this moment? The deliverable exists for them.

This is a load-bearing accuracy tool, not empathy theatre. Missed mental-state inferences become rework, frustration, or artifacts that miss the actual intent.

## Phase 1: Who and What

Use AskUserQuestion to gather context. Ask in batches, not one by one.

**Batch 1 — Identity:**
- "Who is my Boss? (name for Git Discipline references)"
- "What is this project? (one sentence — what problem does it solve?)"

**Batch 2 — Shape:**
- "Who uses this? (audience/users)"
- "What scale are we targeting? (prototype, MVP, production, enterprise)"
- "Single dev or team? (affects code review, conventions depth)"

## Phase 2: Explore Existing Code

If the project already has code, launch a **code-explorer** agent to deeply analyze the codebase:

Spawn the code-explorer agent with this prompt:
"Analyze this project. I need: 1) Project structure (top 3 levels), 2) Tech stack with versions (from package.json, go.mod, requirements.txt, etc.), 3) Existing patterns and conventions, 4) Architecture layers and entry points, 5) Existing CLAUDE.md, README.md, or documentation, 6) Tests, CI/CD, database configs. Provide a concise summary."

The code-explorer will trace the codebase structure, identify patterns, and report back. Use its findings to:
- Inform convention selection in Phase 3
- Pre-fill the tech stack in the PROJECT.md
- Detect existing patterns that should be preserved

Summarize the code-explorer's findings to Boss: "Here's what I found in the codebase: [tech stack, structure, patterns]"

If no code exists, skip to Phase 3.

## Phase 3: Configure Conventions

1. Ask Boss: "Want me to generate project conventions based on the tech stack and what I found in the code?"
2. If yes, write a `project-conventions.md` in `.claude/rules/` tailored to the detected stack and patterns — it auto-loads as a project rule every session

Do not use `misc/archive/` (or `atlas/misc/archive/`) — it holds retired material (old context templates) kept for reference only.

## Phase 4: Configure Workspace

### Update Boss Name
- Edit CLAUDE.md: Replace "Boss Kamil" with "Boss {bossName}" in Git Discipline section

### Configure Repos (multi-repo mode only)
If `repos/` directory exists:
1. List directories in `repos/`
2. For each repo, use AskUserQuestion:
   - "What port does {repo} run on?"
   - "What is the startup command?" (suggest defaults based on detected tech stack)
3. Update `repos/CLAUDE.md` with repo info
4. Update `.claude/commands/start/be-fe.md` with actual commands

### Configure MCP (if .mcp.json exists)
Don't re-ask about Playwright — the scaffolder already asked the browser automation choice (none / Playwright MCP / Playwright CLI) during `new-project`.
- If PostgreSQL MCP is configured, ask: "Update the connection string?"

## Phase 5: Generate Project Context

Write `.claude/rules/PROJECT.md` with this structure:

```markdown
# {Project Name}

## What This Is
{One paragraph — what the project does and why it exists}

## Who Uses It
{Target audience and their needs}

## Scale & Constraints
- **Scale target**: {prototype/MVP/production/enterprise}
- **Team size**: {solo/small team/large team}
- **Key constraints**: {deadlines, tech limitations, business rules}

## Tech Stack
{Detected or declared tech stack with versions}

## Architecture Decisions
{Key decisions made so far — framework choices, database, deployment}

## What's Next
{Current priorities or next features to build}
```

## Phase 6: Summary

Report what was configured:
- Boss: {name}
- Project: {one-liner}
- Tech stack: {detected/declared}
- Conventions activated: {list}
- Repos configured: {list with ports, if applicable}
- Files created/updated: {list}

Remind Boss: "Run `git diff` to review changes. When ready, I'll commit."
abstraction-powerSkill

Activate ATLAS pattern recognition mode. Identify repeated patterns, extract essential characteristics, and create reusable abstractions from concrete examples.

adversarial-reviewSkill

Adversarially review something just built — presume it is broken and find where. Use after implementing a feature, finishing a build, or before shipping, or whenever ATLAS or Boss wants a hostile second opinion on a diff, a running app, or a whole codebase. Spawns clean-context reviewers with opposing lenses (correctness, security, empty-world) and reproduces every finding by running the target before believing it. Distinct from /code-review, which statically reads a diff and never runs the app.

algorithmic-artSkill

Creating algorithmic art using p5.js with seeded randomness and interactive parameter exploration. Use this when users request creating art using code, generative art, algorithmic art, flow fields, or particle systems. Create original algorithmic art rather than copying existing artists' work to avoid copyright violations.

anthropic-product-knowledgeSkill

Stop and consult this skill whenever your response would include specific facts about Anthropic's products. Covers: Claude Code (how to install, Node.js requirements, platform/OS support, MCP server integration, configuration), Claude API (function calling/tool use, batch processing, SDK usage, rate limits, pricing, models, streaming), and Claude.ai (Pro vs Team vs Enterprise plans, feature limits). Trigger this even for coding tasks that use the Anthropic SDK, content creation mentioning Claude capabilities or pricing, or LLM provider comparisons. Any time you would otherwise rely on memory for Anthropic product details, verify here instead — your training data may be outdated or wrong.

canvas-designSkill

Create beautiful visual art in .png and .pdf documents using design philosophy. You should use this skill when the user asks to create a poster, piece of art, design, or other static piece. Create original visual designs, never copying existing artists' work to avoid copyright violations.

change-core-selfSlash Command

Interview Boss about the project, then reason from first principles to design the ideal ATLAS operating identity/system-prompt for it — free to drop KISS/YAGNI/DRY/clean-architecture entirely when the project (and the LLM's own distribution) calls for a different mindset

code-architectSubagent

Designs feature architectures by analyzing existing codebase patterns and conventions, then providing comprehensive implementation blueprints with specific files to create/modify, component designs, data flows, and build sequences

code-explorerSubagent

Deeply analyzes existing codebase features by tracing execution paths, mapping architecture layers, understanding patterns and abstractions, and documenting dependencies to inform new development