[Enhancement] API access for Cowork global and project instruction layers — close automation parity gap

[daskuntal75]

Summary

Claude's instruction system has 7 layers. Five can be updated programmatically (files on disk or existing API). Two — the Cowork project and global instruction layers — require manual copy-paste in the browser UI. This creates a two-tier automation system with no path to close the gap.

Current state

Layer	Auto-updatable?	Method
L3 Code CLAUDE.md (global + project)	✅	File on disk
L5 Memory files	✅	Files on disk
L7 Chat project instructions	✅ (read)	API
L1 Cowork project instructions	❌	UI only
L2 Cowork global instructions	❌	UI only

Real-world impact

An hourly automation pipeline successfully refreshes Code CLAUDE.md and Chat project instructions with current cost/usage data — but Cowork layers must be copy-pasted manually into the browser. The instructions drift and become stale between manual updates.

Proposed API

GET  /v1/cowork/instructions/global
PUT  /v1/cowork/instructions/global        { "content": "..." }
GET  /v1/cowork/projects/{id}/instructions
PUT  /v1/cowork/projects/{id}/instructions { "content": "..." }

Minimum viable version

Read-only GET endpoints would at least enable validation and diff tooling. Full read/write would enable automation parity across all 7 layers.

Reference

Open-source tooling impacted by this gap: https://github.com/daskuntal75/llm-cost-kit
7-layer instruction architecture explanation: https://github.com/daskuntal75/llm-cost-kit/blob/main/docs/responsible-ai-cost-framework.md
Full issue context: https://github.com/daskuntal75/llm-cost-kit/blob/main/docs/anthropic-issues-submit-ready.md