Setup complete

You're ready to prompt well.

PromptWell sits between you and Fable 5.
Messy prompt in. Crisp prompt out. Fewer tokens burned.

4 commands — that's it

# 1. Set up once
npx promptwell init

# 2. Before you talk to Fable 5 — paste the output in instead
npx promptwell crisp "fix the login button not working on mobile"

# 3. After Fable 5 finishes — record what happened
npx promptwell score

# 4. See how you're improving over time
npx promptwell stats

How it works

You Messy prompt "fix the auth bug"
→
Haiku 4.5 crisp() ~$0.001
→
Brainstorm Q&A skip any question
→
Blueprint Crisp prompt score + savings
→
Fable 5 Executes no wasted tokens

The 3 tools

crisp(task, context?) — optimize a prompt before Fable 5 sees it
Call this whenever you're about to hand a vague task to Fable 5. It returns a clean prompt, a disambiguation score, and an estimated token savings. 
// In Claude Code, call it as an MCP tool:
mcp__promptwell__crisp({
  task: "fix the auth bug",
  context: "Next.js app, JWT auth in src/lib/auth.ts"  // optional
})

// Returns:
{
  crisp_prompt: "Goal: Fix JWT token expiry handling...\n\nConstraints:\n- do not change token format\n...",
  blueprint: {
    structural_goal: "Fix JWT token expiry in auth middleware",
    disambiguation_score: 72,
    what_was_vague: ["no file path specified", "unclear which auth system"]
  },
  estimated_token_savings: "~2,880 tokens"
}
0–20Crystal clear — specific paths, measurable outcome 21–50Decent — goal clear, missing some context 51–80Vague — common, fixable patterns 81–100Unusable — Fable 5 will guess major requirements
score(...) — record what Fable 5 accomplished
Call this after Fable 5 finishes. It saves the session to your history so PromptWell can track your improvement over time. 
mcp__promptwell__score({
  original_prompt: "fix the auth bug",
  disambiguation_score: 72,          // from crisp()
  what_was_vague: ["no file path specified"],  // from crisp()
  result_summary: "Fixed JWT expiry, all tests pass",
  tokens_used: 4200                  // optional
})
stats() — see your prompting trends
Shows your average disambiguation score over time, token savings, and the top patterns you repeat so you can fix them. 
mcp__promptwell__stats()

// Example output:
PromptWell Stats
Sessions tracked: 24
Avg disambiguation score: 48/100 (lower = crisper)
7-day avg:  41/100
30-day avg: 56/100
Estimated tokens saved: ~38,400

Top patterns to fix:
  - no file path specified (seen 14x)
  - no success criteria (seen 9x)
  - unclear which system (seen 6x)

Using inside Claude Code

MCP mode — no terminal needed — Claude Code calls PromptWell for you
Once npx promptwell init runs, PromptWell is registered as an MCP server. Claude Code can call it automatically — you never need to leave the chat.
# Trigger it explicitly:
"crisp this before we start: fix the login button"
"before you begin, run this through promptwell: [task]"
"optimize my prompt: [task]"

# Or let SKILL.md auto-trigger when your prompt is vague:
"fix the login bug" → Claude detects vagueness → calls crisp() automatically
Important: Claude Code can't ask you interactive questions the way the terminal can. Pack your context into the message upfront — file path, expected vs actual behavior, constraints. The more you give, the better the output. 
// Rich context = much better crisp prompt:
"crisp this: fix the login button in src/components/LoginButton.tsx —
 it shows on iOS Safari but tap does nothing. Don't change the API contract."

// Claude Code routes it as:
mcp__promptwell__crisp({
  task: "fix the login button",
  context: "src/components/LoginButton.tsx, iOS Safari, tap has no effect"
})
Tip: The best time to call crisp() is right before you'd normally type a long, multi-sentence prompt to Fable 5. If you can describe the task in one crisp sentence with a file path and a success criterion, you probably don't need it. If you can't — you do.