I was burning through Claude limit in 2 hours - built a tool that cuts it 15x-22x

[anshmajumdar121]

Like a lot of you this week, I kept hitting the cap faster than ever. Tried all the usual stuff - /clear, shorter prompts, fresh sessions. Helped a bit but did not fix the root cause.

The real problem: every turn, Claude re-reads your entire project structure even when it only needs 3 files. On a medium codebase that is 15,000-30,000 tokens of dead weight per message.

My fix: give Claude a pre-built structural index (CONTEXT_MANIFEST.md) that maps every file, its imports, and its blast-radius. Claude reads this instead of scanning directories.

Results

Task	Before	After	Reduction
Code review (3 files)	~18,000 tokens	~1,200	15x
Debug a function	~8,000	~400	20x
Plan a feature	~35,000	~1,800	19x
Full monorepo analysis	~80,000	~3,500	22x

How it works

1. Run a local Python script (zero deps, Python 3.7+) that indexes your project into CONTEXT_MANIFEST.md
2. Upload to a Claude Project as Knowledge
3. Custom instructions force Claude to read manifest first, fetch max 3 files per turn, output compressed

No API hacks. No leaked internals. Just Claude Projects + Custom Instructions used differently.

Quick start

git clone https://github.com/anshmajumdar121/context-optimizer.git
cd context-optimizer
chmod +x scripts/install.sh
./scripts/install.sh /path/to/your/project
python3 tools/context_mapper.py /path/to/your/project

Installer auto-detects Cursor, Windsurf, Claude Code, Cline, and Copilot.

Curious if others have tried the manifest approach or found something that works better?

https://github.com/anshmajumdar121/context-optimizer