[FEATURE] Zero-Cost Intra-Project Operations & Intelligent Context Management for Claude.ai Projects

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Claude.ai Projects treat intra-project operations (file loading, chat search, reasoning about project content) identically to external data requests, consuming user tokens for accessing their own workspace content. Additionally, RAG search activates at ~13 files / 2% project capacity, far below the documented threshold of "approaching context window limit." This makes complex multi-file projects unusable. A 700+ chapter creative writing project with 15 reference files totaling 73K tokens (35% of 200K window) should load directly — instead, the system fragments files through search, causing hallucinations, wrong names, invented data, and broken consistency. Users cannot expand their projects without worsening this behavior. Related bug: #25759

Proposed Solution

Raise or remove RAG threshold — 73K tokens in 200K window must load directly, not through search 2. Zero-cost intra-project file access — loading your own project files should not consume creative token budget 3. Tiered optimization for paying subscribers — higher tier = greater reduction on internal operations 4. Lazy file loading — three states: Active (in context), Dormant (zero cost until triggered), Archived (search only) 5. Replace-not-accumulate — chapter revisions overwrite instead of stacking versions 6. Claude write access to project files with user permission for world state updates 7. Per-file token counts visible in project UI 8. Lexical control module — word blacklists and auto-substitution for genre-specific writing 9. Persistent world state tracking between sessions

Alternative Solutions

Currently I limit project to 10 files, write reference files in English instead of Russian to save tokens, manually count tokens using external tokenizer, and rewrite chapters when RAG causes hallucinations. None of these are acceptable long-term solutions for a paying subscriber.

Priority

Critical - Blocking my work

Feature Category

Other

Use Case Example

Anthropic_Feature_Request_Complete x.md
Anthropic_Feature_Request_Zero_Cost_Intra_Project_Operations.docx

1. I run a 550+ chapter alternate-universe Game of Thrones project with 15 interconnected reference files covering character genealogies, economics, military structures, geography, and political appointments — totaling ~73K tokens 2. Two days ago (Feb 12, 2026) all 15 files loaded directly into context — perfect consistency 3. Today (Feb 14, 2026) at 13 files, system switches to RAG at 2% capacity with message "To save space in chats, Claude will look up specific information as needed" 4. RAG retrieves fragments, misses cross-file connections — Claude hallucinates character names, invents locations, fabricates economic data that contradicts my files 5. I now must cut 5-6K tokens of character development from files just to stay under an arbitrary file count threshold 6. I write files in English instead of Russian to save tokens, manually count tokens via external tokenizer 7. I pay $200+/month and spend time debugging the system instead of writing

Additional Context

Related bug report: #25759 — RAG threshold activates at 2% capacity, threshold based on file count not token size. Screenshots attached showing 10 files/1% (no warning, direct loading) vs 13 files/2% (RAG warning). Detailed feature request documents submitted to Anthropic support. Max subscriber ($200/month) from Ukraine.

<img width="645" height="1398" alt="Image" src="https://github.com/user-attachments/assets/fa48b97d-89e6-4915-b5b6-2d60854f192e" />
<img width="645" height="1398" alt="Image" src="https://github.com/user-attachments/assets/18ef2cfd-e2fb-4d08-b828-0fbcada4a9a9" />