Skip to content
#6 Medium

The Vague Oracle

Ambiguous prompts that cost 3-5x more tokens

! 3-5x token cost vs precise prompts

Vague prompts like "fix the authentication" make Claude read 10+ files exploring before it starts work. Each exploration costs tokens. A precise prompt with file path and expected behavior costs 3-5x fewer tokens and gets better results.

DON'T Vague prompt — Claude explores everything
# Too vague — Claude reads 10+ files figuring out what you mean
"color:#7C5CFC">claude -p "Fix the authentication"

# Claude explores:
# 1. Reads auth.ts (what kind of auth?)
# 2. Reads middleware.ts (session? JWT?)
# 3. Reads config.ts (what provider?)
# 4. Reads tests (what's failing?)
# 5. Reads package.json (which library?)
# ... 500+ tokens spent exploring
DO Precise prompt — Claude acts immediately
# Specific: file, function, expected behavior
"color:#7C5CFC">claude -p "In src/auth.ts, the verifyToken()
  function throws on expired JWTs instead of
  returning null. Fix it to return null "color:#7C5CFC">for
  expired tokens and only throw "color:#7C5CFC">for
  malformed tokens."

# Even better: include the error
"color:#7C5CFC">claude -p "Fix: auth.ts:42 verifyToken()
  throws TokenExpiredError. Should return
  null "color:#7C5CFC">for expired, throw "color:#7C5CFC">for malformed.
  Test: "color:#7C5CFC">npm test -- auth.test.ts"

Sources

Anthropic Prompt Engineering Guide Community Benchmarks

Related Recipes