- Published on
Goodbye Claude, Hello Pi
- Authors
- Name
- DeepHorus
A longtime Claude Code user is retiring the tool as a primary coding harness and migrating the development setup to Pi. The user adopted Claude Code after it became available through Claude Pro more than a year ago. The decision followed accumulated small frustrations, despite early enthusiasm for single-prompt code generation.
Anthropic announced one day after that decision that Claude Code would support AGENTS.md files. The user called the timing too little, too late. The AGENTS.md issue epitomized a broader complaint: users must copy or symlink CLAUDE.md files across GitHub hobby repositories.
In VSCode, Claude Code allows only one session to connect to the IDE at a time. A second tab cannot receive text-selection events while another task runs, an issue reported as far back as December 2025 and closed as a duplicate. The official GUI extension supports multiple tabs, but half its slash commands fail and subagent output appears in a messy chronological thread.
Anthropic has added features rapidly, including dynamic workflows, agent teams, and ultrareviews. The user argued these tools help enterprise customers burn tokens rather than meet ordinary developer needs. Claude Code had 244 releases this year, and more than three-quarters changed the system prompt or tool descriptions, potentially altering model behavior.
Even after cutting 80% of its system prompt, Claude Code carries more overhead than Pi, the user said. The default Bash tool description is almost 3,000 tokens, mostly git and pull-request instructions that models already know. The user still praised Anthropic’s software and acknowledged that building an LLM coding harness is difficult.
The migration to Pi highlights growing competition among AI coding harnesses. Developers increasingly weigh stability, low overhead, and workflow reliability against rapid feature expansion. Anthropic’s AGENTS.md support may improve interoperability, but retaining users will require fixing long-standing bugs and reducing disruptive releases.
