promptkit-claude 0.1.0

Token-efficient prompting and agentic workflows for Claude

promptkit-claude

Token-efficient prompting and agentic workflows for Claude.

Install

pip install promptkit-claude

Quick Start — Offline (no API key needed)

from claude_kit.compress import compress
from claude_kit.templates import summarize

# Remove filler phrases from any prompt (no API key needed)
prompt = compress("Please carefully analyze the following and provide a detailed summary")
# → "analyze the following and provide a summary"

# Build a token-efficient prompt
prompt = summarize(document, format="bullet", max_bullets=5)

Quick Start — With API Key

from claude_kit.cache import CachedClient

client = CachedClient()

# NOTE: First call warms the cache — expect 0% hit rate.
# Subsequent calls with the same system prompt hit the cache.
response = client.message(
    system="You are a helpful assistant with a long system prompt...",
    user="Hello!",
    cache_system=True,
)

print(client.cache_stats())
# → {"cache_hit_rate": "74%", "tokens_saved": 1840, "savings_usd": "$0.0023"}

Session Cost Tracking

from claude_kit.tracker import Tracker

with Tracker() as tracker:
    response = tracker.client.messages.create(
        model="claude-sonnet-4-5",
        max_tokens=1024,
        messages=[{"role": "user", "content": "Summarize quantum computing."}],
    )

print(tracker.summary())
# → {"input_tokens": 12, "output_tokens": 340, "total_cost_usd": "$0.0014"}

Agent Composition

from claude_kit.agent import Chain, Retry

# Chain: pipe output of one prompt into the next
result = Chain([
    "Summarize this document: {input}",
    "Extract action items from: {input}",
]).run(input=document)

# Retry: retry on failure with exponential backoff
safe_call = Retry(max_attempts=3).wrap(client.message)
response = safe_call(user="Tell me a joke.")

Async Support

from claude_kit.cache import AsyncCachedClient

client = AsyncCachedClient()

async def main():
    response = await client.message(
        system="You are a concise assistant.",
        user="What is 2 + 2?",
        cache_system=True,
    )
    return response

CLI

promptkit-claude compress "Please carefully explain this in detail"
# → "explain this in detail"

Modules

Module	What it does	Needs API key?
compress	Removes filler phrases from prompts	No
templates	Token-efficient prompt templates	No
tracker	Session cost tracking	Yes (wraps SDK)
cache	CachedClient with cache_control	Yes
agent	Chain / Retry / Fallback / Parallel	Yes

Versioning

We follow semver. Patch: bug fixes. Minor: new APIs, backwards compatible. Major: breaking changes with migration notes in CHANGELOG.

v0.2 Roadmap

• Fallback and Parallel agent primitives with full test coverage
• cache_tools support in CachedClient for tool-use payloads
• templates.classify() and templates.extract() for common extraction tasks
• Token budget warnings and hard caps via Tracker
• Rich CLI output with cost breakdown per call
• LangChain and LlamaIndex adapter shims