@MaximeRivest: Pi coding agent is very often the best for its price. Pi's system prompt is very short and has only 4 tools. My Pi syst…
Summary
Maxime Rivest shares tips for optimizing Pi coding agent usage with custom system prompts, and Matei Zaharia discusses surprising results from benchmarking coding agents at Databricks.
View Cached Full Text
Cached at: 07/09/26, 07:51 PM
Pi coding agent is very often the best for its price.
Pi’s system prompt is very short and has only 4 tools.
My Pi system prompt is even more minimalistic and I get better & cheaper results (according to my guts; sorry i don’t have better proof, yet!)
how?
I have a custom extension that let’s me change the system prompt easily and I now have 25 modes (prewritten specialized system prompts).
The models are better at smaller context, everyone accept that as fact by now. I, also, feel that everyone would accept as fact that a model is better with a specialized system prompt. Yet, no one does it for coding agents! Pi agents are told that they are coding agents in Pi, I prefer them to think they are maintainer of library X and have access to some tools. Or that they are Documentation writer, etc.
LLMs are trained to be extremely ‘responsive’ and ‘deligent’ in following system prompts over user prompts. We must take advantage of that!
I often switch system prompt mid-conversation:
oh not my prompt cache!
its actually not bad and in my experience it is totally worth it! Thanks to the gain in performance for the next turn the model use much less tool calls and succeed faster, offsetting the cost of not being prompt caching friendly.
In short, I recommend pi and i recommend making yourself a way to change the system prompt.
here is the vibed repo, intended as inspiration.
I included some of my modes. my favorite mode are the stackoverflow and surprise-reporter.
https://github.com/MaximeRivest/pi-modes…
i switch the system prompt often for turns lower down i nthe conversation, system prompt instead of user prompt because it stirs the model more forcefully. replacing pi prompt with mine instead of appending to not have the ‘confusing’:
you are a pi coding assistant. [append] you are a stackoverflow response simulator.
i also toggle in our out tools description, in somecase making the model think no tool exists (so it just responds).
true! I never thought it was of much interest. let me push it.
here is the vibed repo, intended as inspiration.
I included some of my modes. my favorite mode are the stackoverflow and surprise-reporter.
https://github.com/MaximeRivest/pi-modes…
Similar Articles
@pidotdev: Read the full blog post here
A blog post highlighting Pi, a minimal coding agent harness, arguing that its simplicity yields better performance and lower cost compared to more complex tools, supported by Databricks benchmark results and Shopify's Pi Autoresearch case study.
@binghe: I've recently become a heavy user of Pi Agent... The more I use it, the more I like it. Pi can chain multiple skills to complete complex tasks, and the local Agent has more practical value! Pi Agent's system prompt is less than 1500 characters, only about 1/13 of Claude Code's...
The user shares their in-depth experience with Pi Agent, noting that its prompts are short, responses fast, and token consumption low, outperforming Claude Code.
According to DataBricks, pi-coding-agent is ~2x cheaper than CC/Codex, GLM 5.2 on par with Opus 4.8 high
Databricks benchmarks show pi-coding-agent is up to 2x cheaper than CC/Codex with higher pass rates, and GLM 5.2 performs on par with Opus 4.8 for coding tasks.
@rohanpaul_ai: Really useful revelation on prompting coding agent in this paper. Your coding-agent prompt may be spending compute on w…
The paper reveals that certain prompt instructions in coding agents can lead to redundant work without improving success rates, and recommends using bounded instructions to minimize waste.
Pi Coding Agent
Pi Coding Agent is a customizable harness for building coding agents, now available on Product Hunt.