DEV Community

Abkari Mohammed Sayeem
Abkari Mohammed Sayeem

Posted on

Context Engineering > Prompt Engineering: I built a Python lib that does it automatically

#ai #python #llm #opensource

Stop Copy-Pasting Files Into ChatGPT

Every developer using AI coding tools hits the same wall:

  • Which files do I paste into Claude?
  • Token limit exceeded... again
  • Bad answer because I included the wrong context
  • 10 minutes wasted before even asking the question

I got frustrated with this daily and built ctxeng — a Python library that solves it automatically.

What Is Context Engineering?

Context engineering is the practice of carefully constructing what goes into an LLM's context window. It's more impactful than prompt engineering because:

The quality of your LLM's output depends almost entirely on what you put in the context window — not how you phrase the question.

The problem is doing it manually is painful. ctxeng automates it.

How ctxeng Works 

from ctxeng import ContextEngine

engine = ContextEngine(root=".", model="claude-sonnet-4")
ctx = engine.build("Fix the authentication bug in the login flow")

print(ctx.summary())
# Context summary (12,340 tokens / 197,440 budget):
#   Included : 8 files
#   Skipped  : 23 files (over budget)
#   [████████  ] 0.84  src/auth/login.py
#   [███████   ] 0.71  src/auth/middleware.py
#   [█████     ] 0.53  src/models/user.py
#   [████      ] 0.41  tests/test_auth.py

print(ctx.to_string())  # paste this directly into Claude
Enter fullscreen mode Exit fullscreen mode

It scans your entire codebase, scores every file for relevance, and fits the best ones into your model's token window.

The Scoring Algorithm

Every file gets a relevance score from 0 → 1 using four signals:

1. Keyword Overlap

How many query terms appear in the file content. Simple but surprisingly effective.

2. Python AST Analysis

ctxeng actually parses your Python files and extracts class names, function names, and import names — then matches them against your query. So if you ask about "authentication", files containing class AuthService or def authenticate() score higher.

3. Git Recency

Files touched in recent commits score higher. The logic: if you're asking about a bug, the files you recently changed are probably related.

4. File Path Matching

Filename and directory names matched against query tokens. src/auth/login.py scores higher for auth-related queries.

Fluent Builder API

For more control, use the chainable builder:

from ctxeng import ContextBuilder

ctx = (
    ContextBuilder(root=".")
    .for_model("gpt-4o")
    .only("**/*.py")
    .exclude("tests/**", "migrations/**")
    .from_git_diff()           # only changed files
    .with_system("You are a senior Python engineer.")
    .build("Refactor the payment module to use async/await")
)

print(ctx.to_string("markdown"))
Enter fullscreen mode Exit fullscreen mode

CLI 

# Build context for a query
ctxeng build "Fix the auth bug"

# Only git-changed files
ctxeng build "Review my changes" --git-diff

# Save to file
ctxeng build "Explain the payment flow" --output context.md

# Project stats
ctxeng info
Enter fullscreen mode Exit fullscreen mode

Smart Token Truncation

When a file doesn't fully fit the budget, ctxeng doesn't just cut it off. It keeps:

  • The head (imports, class definitions, docstrings)
  • The tail (most recent changes)
  • Drops the middle

Both ends are high-signal. The middle is usually implementation details.

One-Line LLM Integration 

from ctxeng import ContextEngine
from ctxeng.integrations import ask_claude

engine = ContextEngine(".", model="claude-sonnet-4")
ctx = engine.build("Why is the test_login test failing?")

response = ask_claude(ctx)
print(response)
Enter fullscreen mode Exit fullscreen mode

Works With Every Model

Model Window Auto-detected
claude-sonnet-4 200K ✓
gpt-4o 128K ✓
gemini-1.5-pro 1M ✓
Local models custom set manually

For local models just set the budget manually:

ctx = (
    ContextBuilder(".")
    .with_budget(total=32_768)
    .build("Refactor this module")
)
Enter fullscreen mode Exit fullscreen mode

Installation 

pip install ctxeng

# With accurate token counting
pip install "ctxeng[tiktoken]"

# With Claude integration
pip install "ctxeng[anthropic]"

# Everything
pip install "ctxeng[all]"
Enter fullscreen mode Exit fullscreen mode

Zero required dependencies by default. MIT license.

What's Next

  • Semantic similarity scoring (embedding model)
  • ctxeng watch — auto-rebuild on file changes
  • VSCode extension
  • Import graph analysis

Try It

⭐ GitHub: https://github.com/Sayeem3051/python-context-engineer

Would love feedback — especially on the scoring algorithm. What signals would make this more useful for your workflow?

Top comments (0)