Docs as Source of Truth, Models as Suspect

Models lie about themselves

Every Claude model has a knowledge cutoff date. Ask the model what the newest model is and it answers from training data — sometimes years stale. The same goes for SDK names, function signatures, default options, and pricing. Models speak with confidence about a world that has moved on.

Trust order

For SDK or model questions, the trust order is: (1) docs you can fetch right now, (2) the source code installed in your environment, (3) the model's own answer. Reverse this order and you will ship code against an SDK shape that has been deprecated, or pin a model id that no longer exists.

How cwkPippa works around this

cwkPippa's docs index lives in the Obsidian vault as canonical text. When Claude-Pippa needs to answer an SDK question, the system prompt already carries the latest fingerprints of backend/adapters/claude.py and the relevant Anthropic docs. The model is a reasoner over current text, not a knowledge base.

Principle: Trust the docs you fetched five minutes ago. Distrust the model's recollection of last year.

Code

Verify SDK shape from the installed package·bash

# What is actually installed in this venv?
pip show anthropic | head -n 5
python -c "from anthropic import Anthropic; import anthropic; print(anthropic.__version__)"

# What does this version expose?
python -c "from anthropic.types import Message; print([f for f in Message.model_fields])"

Pull docs into the model's context, don't trust memory·python

import httpx
from anthropic import Anthropic

docs = httpx.get("https://docs.claude.com/llms.txt").text  # Anthropic publishes a flat llms.txt index.

client = Anthropic()
resp = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=512,
    system="Answer using ONLY the docs provided. Quote the section you used.",
    messages=[
        {"role": "user", "content": f"<docs>\n{docs}\n</docs>\n\nWhat package is the renamed Agent SDK on PyPI?"},
    ],
)
print(resp.content[0].text)

Docs as Source of Truth, Models as Suspect

Models lie about themselves

Trust order

How cwkPippa works around this

Code

External links

Exercise

Progress

Comments 0