pi-model-system-prompt

Lets different large models use different system prompts. When you switch models, the system prompt specific to that model is appended to the context, keyed on the currently active model id.

中文

Why this extension

Large models need more than different capabilities — they need different harnesses.

The level of the harness matters enormously to how well a large model performs, which is by now a given across the industry.

This shows up in agents such as Claude Code, Codex, and the pi agent, where each loads a system prompt tailored to a specific model at startup.

In the course of using the pi agent we mix and match large models from several different sources. That kind of use is exactly the charm of a coding agent as free-form as pi.

So when switching between models inside the pi agent, the ideal is that each model gets to use its own independent system prompt — but the pi agent itself has no such capability.

This extension provides exactly that.

Install

pi install npm:pi-model-system-prompt

Restart Pi after installing.

Quick start

It takes effect as soon as it is installed. The extension picks the Markdown file in ~/.pi/agent/model-system-prompt/ whose name is the longest prefix of the current model id, and appends its contents to the pi agent's system prompt. For model glm-5.3-flash that is glm-5.3-flash.md if it exists, otherwise glm-5.3.md, otherwise glm.md. Case is ignored, and a model with no matching file is untouched.

To write your own system prompt for a specific model, add a Markdown file in ~/.pi/agent/model-system-prompt/ named after that model — or after the family it belongs to, which then serves every model id starting with that name — and write the prompt in it. It is loaded automatically when you open pi.

The extension ships well-tuned preset prompts. Note that open-source models come in different quantizations and fine-tunes, so every preset prompt is tuned against its own official model, and no guarantee is made for unofficial builds. You are encouraged to tune the system prompt yourself to match the model you actually run.

Models with preset system prompts

• MiniMax-M3
• MiniMax-M3.1-Flash-Preview
• claude-sonnet-5, claude-haiku-5 (Claude Code's full system prompt)
• claude-opus-5, claude-fable-5 (Claude Code's shorter system prompt)
• deepseek-flash
• glm-5.3-flash
• glm-5.3
• gpt-5.6 (gpt-5.6-luna, gpt-5.6-sol, gpt-5.6-terra — Codex ships one prompt for all three)
• gpt-6-luna, gpt-6-sol, gpt-6-astra, gpt-6.1-sol (Codex rewrites the prompt per tier, so one file each)
• kimi (kimi-k3 and the rest of the family)

Matching and override rules

Which file wins is decided by its name

1. The longest prefix wins. When several files match the same model id, the one with the longest name takes effect. For id glm-5.3-flash, with glm-5.3-flash.md, glm-5.3.md and glm.md all present, only glm-5.3-flash.md applies.
2. A file named exactly after the id always wins.
3. Case is ignored.
4. No wildcards, literal prefixes only. glm.md also matches an unrelated id like glmish-2. So do not pick too short a name: g.md swallows every id starting with g. The prefix counts from the start of the id, so a file named claude-opus-5.md does not match a provider-prefixed id such as anthropic.claude-opus-5.

Turning a model off

• Recommended: empty the file that would win. The file exists but is empty = the model gets no dedicated prompt, and it does not fall back to a shorter file of the same family.
• Deleting a file is not turning it off. It only counts while no shorter name still matches: delete glm-5.3.md and glm-5.3 is still served by glm.md.
• To turn off a whole family, emptying the family file only covers ids with no more specific file — empty glm.md and glm-4 gets nothing, but as long as glm-5.3.md is there, glm-5.3-flash still goes through glm-5.3.md. To really kill a family, empty every relevant glm*.md.

Where the preset prompts come from

Each comes from the official coding agent tool of that model itself

minimax from minimax code

glm from zcode

deepseek from deepseek harness

kimi from kimi code (one harness for the whole model family, so one family file)

claude from claude code (2.1.285, stable channel — it ships a full and a shorter system prompt and picks by model; both are mirrored, one family file each)

gpt from codex (openai/codex, tag rust-v0.161.0; the text for a model id is the instructions_template field in codex-rs/models-manager/models.json, and the shipped @openai/codex@0.161.0 binary carries the same bytes. Codex gives one prompt to the whole gpt-5.6 family and a rewritten one to each gpt-6 tier, so 5.6 is a single family file while each gpt-6 tier has its own — a later generation ships its own text and needs its own file.)

New preset model prompts are welcome

To verify generality, the source must be stated clearly and the prompt must be tuned against the official model

What happens on first run

☕ Support the project

If you find this project useful, consider supporting its development.