steelprompt

---

The problem

Claude is only as good as the prompts you give it. Most prompts are missing role assignment, structured context, explicit constraints, output format specification, and examples — all things Anthropic's own guidelines say dramatically improve response quality.

You could spend 10 minutes engineering every prompt manually. Or use steelprompt — automatic in Claude Code, on demand in Claude.ai.

---

How it works

steelprompt runs on two surfaces:

• Claude Code (CLI) — a UserPromptSubmit hook intercepts every prompt automatically before Claude processes it.
• Claude.ai (web) — paste prompts/steelprompt-web.md into Custom Instructions once, then use /sp "prompt" to engineer any prompt on demand.

In both cases, the same 3-tier decision protocol applies:

Your prompt
    │
    ▼
┌─────────────────────────────────────────────────┐
│  TIER 1 — BYPASS                                │
│  Slash commands · < 5 words · atomic tasks      │
│  → Pass through unchanged                       │
└─────────────────────────────────────────────────┘
    │ not bypassed
    ▼
┌─────────────────────────────────────────────────┐
│  TIER 2 — ASK                                   │
│  Critical information missing?                  │
│  → AskUserQuestion (1–2 targeted questions)     │
└─────────────────────────────────────────────────┘
    │ prompt is clear
    ▼
┌─────────────────────────────────────────────────┐
│  TIER 3 — APPLY FRAMEWORK           ← default   │
│  Restructure using 7 Anthropic principles       │
│  silently, before Claude responds               │
└─────────────────────────────────────────────────┘

No external API calls. Claude restructures inline during its normal inference — zero added latency, zero extra tokens billed.

---

Demo

---

Tier 3 in action

This is the core of steelprompt. Every clear, non-atomic prompt goes through it automatically.

You type:

add rate limiting to the auth endpoints

What steelprompt injects internally (visible in preview mode before execution):

You are a senior backend security engineer.

<context>
Project: Express.js API with JWT authentication
Endpoints affected: POST /auth/login, POST /auth/refresh, POST /auth/register
Environment: Node.js 20, existing middleware stack in src/middleware/
Known constraints from conversation: Redis already in use for session storage
</context>

<task>
Implement rate limiting on authentication endpoints:
1. Add per-IP rate limiting: max 5 requests/minute on /login and /register
2. Add per-user rate limiting: max 20 requests/hour on /refresh
3. Return 429 with Retry-After header on limit exceeded
4. Do not affect non-auth endpoints
</task>

<constraints>
- Reuse existing Redis connection, do not introduce a new client
- Do not modify JWT validation logic
- Rate limit state must survive server restarts
- Follow existing middleware patterns in src/middleware/
</constraints>

<output_format>
1. New middleware file with implementation
2. Diff showing changes to auth routes
3. Two Jest test cases: limit hit + normal flow
</output_format>

Think through this step by step before answering.

What Claude produces: a complete, production-ready implementation with tests — not a generic overview of rate limiting options.

---

Advanced Patterns

steelprompt extends the 7 core principles with 5 context-specific patterns from the full Anthropic documentation — applied automatically when the prompt signals them:

Chain detection

When a task spans multiple sequential operations, steelprompt detects the chain and shows a plan before executing.

You type:

refactor auth.py, add tests, update the documentation