Goodbye Claude, hello Pi

Nov 8, 2026•karagenit•View Original

TL;DR Highlight

Frequent system prompt changes and token overhead pushed one dev from Claude Code to Pi.

Who Should Read

A developer frustrated by Claude Code's unstable updates and token overhead who is looking at other coding harnesses.

Core Mechanics

Claude Code read only CLAUDE.md, not AGENTS.md
- The author disliked having to copy or symlink between the two formats.
- Anthropic announced AGENTS.md support a day after the author decided to switch.
VS Code integration connects to only one session
- A second tab cannot receive text selection events from the IDE.
- The official GUI extension lacks half the slash commands and a custom status line.
244 releases this year, over 3/4 touched prompts
- There were 244 Claude Code releases this year at the time of writing.
- Over three quarters changed the system prompt or tool descriptions.
- Such changes can substantially alter LLM behavior, especially when they conflict with user prompts or repo conventions.
Bash tool description alone is almost 3,000 tokens
- Most of it covers git usage and opening GitHub PRs.
- The author says models already know this without detailed instructions.
- Even after cutting 80% of the system prompt, overhead is still far above Pi's.
Features outpace product direction
- The author sees dynamic workflows, agent teams, and ultrareviews as ways for enterprises to burn more tokens.
- The author feels nobody is empowered to say no.
Pi is stable and easy to extend
- The author praises its stability, minimalism, and customizability.
- They wrote a multi-tab text-selection extension in about 20 minutes after installing Pi.

Community Reactions

Opinion: don't use Pi with an Anthropic subscription
- A commenter warned of possible bans for ToS violations.
- They argue Claude Code's headers, compaction, and caching are tuned for Anthropic models.
Experience: 5x token burn in non-native harnesses
- A commenter found third-party harnesses sometimes burned 5x the tokens with worse results.
- They now call other harnesses' fresh sessions via MCP instead of calling models directly.
Experience: happy after moving to Pi
- One commenter used Pi daily for over a month and wrote an extension that rewrites complex bash calls into plain English.
- Another moved to OMP, a Pi fork, and reported less frustration and freedom to switch models.
Pushback on the post's logic
- Commenters noted Claude already reads AGENTS.md.
- Others said Pi also ships fast, lacks an auto mode that blocks dangerous actions, and has weak subagent support.

How to Apply

Measure your agent's prompt overhead today
- Check the token count of your harness's system prompt and tool descriptions.
- Look for git or PR instructions the model already knows.
Re-test your rules file after each update
- Run the same task after a Claude Code update to see if CLAUDE.md rules still hold.
- Rewrite any rule that slips to be more specific.
Trial Pi on a side project first
- Install Pi on a small project instead of your main repo and use it for a week.
- Decide whether the extensions you need are worth building yourself.
Check subscription terms and API costs first
- Before attaching Anthropic models to another harness, confirm the terms and API pricing.
- If risky, try Pi with another provider's model.

Try It

copy & run

# Prompt to paste into Claude Code or ChatGPT (simplified check based on the post's claims)

Below are the system prompt and tool descriptions of my coding agent.
{{system prompt and tool descriptions}}

Review them using these criteria.
1. Parts that explain general knowledge the model likely already has (git, writing PRs, etc.)
2. Instructions that may conflict with my repo rules ({{my rules file contents}})
3. Parts that can be trimmed with little impact, and rough token savings

Next step: run the same task with the trimmed version and compare results.

Terminology

harnessA tool that wraps an LLM to edit files, run commands, and manage context. Like a car body that makes the engine (the model) drivable.

PiAn open-source coding harness the author switched to. It's minimal by default and gains features via extensions.

AGENTS.mdA guidance file telling agents about repo rules. Claude Code previously used CLAUDE.md instead.

system promptDefault instructions preloaded before user input. Changing it can change results for the same request.

Related Resources

Goodbye Claude, hello Pi

Nov 8, 2026•karagenit•View Original

TL;DR Highlight

Frequent system prompt changes and token overhead pushed one dev from Claude Code to Pi.

Who Should Read

A developer frustrated by Claude Code's unstable updates and token overhead who is looking at other coding harnesses.

Core Mechanics

Claude Code read only CLAUDE.md, not AGENTS.md
- The author disliked having to copy or symlink between the two formats.
- Anthropic announced AGENTS.md support a day after the author decided to switch.
VS Code integration connects to only one session
- A second tab cannot receive text selection events from the IDE.
- The official GUI extension lacks half the slash commands and a custom status line.
244 releases this year, over 3/4 touched prompts
- There were 244 Claude Code releases this year at the time of writing.
- Over three quarters changed the system prompt or tool descriptions.
- Such changes can substantially alter LLM behavior, especially when they conflict with user prompts or repo conventions.
Bash tool description alone is almost 3,000 tokens
- Most of it covers git usage and opening GitHub PRs.
- The author says models already know this without detailed instructions.
- Even after cutting 80% of the system prompt, overhead is still far above Pi's.
Features outpace product direction
- The author sees dynamic workflows, agent teams, and ultrareviews as ways for enterprises to burn more tokens.
- The author feels nobody is empowered to say no.
Pi is stable and easy to extend
- The author praises its stability, minimalism, and customizability.
- They wrote a multi-tab text-selection extension in about 20 minutes after installing Pi.

Community Reactions

Opinion: don't use Pi with an Anthropic subscription
- A commenter warned of possible bans for ToS violations.
- They argue Claude Code's headers, compaction, and caching are tuned for Anthropic models.
Experience: 5x token burn in non-native harnesses
- A commenter found third-party harnesses sometimes burned 5x the tokens with worse results.
- They now call other harnesses' fresh sessions via MCP instead of calling models directly.
Experience: happy after moving to Pi
- One commenter used Pi daily for over a month and wrote an extension that rewrites complex bash calls into plain English.
- Another moved to OMP, a Pi fork, and reported less frustration and freedom to switch models.
Pushback on the post's logic
- Commenters noted Claude already reads AGENTS.md.
- Others said Pi also ships fast, lacks an auto mode that blocks dangerous actions, and has weak subagent support.

How to Apply

Measure your agent's prompt overhead today
- Check the token count of your harness's system prompt and tool descriptions.
- Look for git or PR instructions the model already knows.
Re-test your rules file after each update
- Run the same task after a Claude Code update to see if CLAUDE.md rules still hold.
- Rewrite any rule that slips to be more specific.
Trial Pi on a side project first
- Install Pi on a small project instead of your main repo and use it for a week.
- Decide whether the extensions you need are worth building yourself.
Check subscription terms and API costs first
- Before attaching Anthropic models to another harness, confirm the terms and API pricing.
- If risky, try Pi with another provider's model.

Try It

copy & run

# Prompt to paste into Claude Code or ChatGPT (simplified check based on the post's claims)

Below are the system prompt and tool descriptions of my coding agent.
{{system prompt and tool descriptions}}

Review them using these criteria.
1. Parts that explain general knowledge the model likely already has (git, writing PRs, etc.)
2. Instructions that may conflict with my repo rules ({{my rules file contents}})
3. Parts that can be trimmed with little impact, and rough token savings

Next step: run the same task with the trimmed version and compare results.

Terminology

harnessA tool that wraps an LLM to edit files, run commands, and manage context. Like a car body that makes the engine (the model) drivable.

PiAn open-source coding harness the author switched to. It's minimal by default and gains features via extensions.

AGENTS.mdA guidance file telling agents about repo rules. Claude Code previously used CLAUDE.md instead.

system promptDefault instructions preloaded before user input. Changing it can change results for the same request.

Goodbye Claude, hello Pi

TL;DR Highlight

Who Should Read

Core Mechanics

Community Reactions

How to Apply

Try It

Terminology

Related Papers

Related Resources

Goodbye Claude, hello Pi

TL;DR Highlight

Who Should Read

Core Mechanics

Community Reactions

How to Apply

Try It

Terminology

Related Papers

Related Resources