EleosWrite the instructions your AI runs under
Add a rule on top, or replace the system prompt itself. Every new chat follows it — and the difference shows up in answers, not just in tone.
The two layers
| Layer | What it does |
|---|---|
| System prompt | Replaces Eleos's built-in base prompt. Write your own, or start from one in the shared catalogue. |
| Rules | Added on top of whichever system prompt is in use — "answer in bullet points", "explain like I'm new to this". |
Adding a rule is a preference. Replacing the system prompt changes the instruction set the model runs under. Most consumer AI apps let you add to their prompt; few let you replace it.
Both reach a Claude account, a key-based provider on the same wire, and a model on your own PC. Chats on ChatGPT, Grok, Cursor or OpenCode do not take either: those agents carry their own harness and the protocol has no field for one.
Where each app keeps it
| App | Where the instruction lives |
|---|---|
| Eleos | Settings → Rules: a system prompt box and a rules box, both editable, both applied to every new chat |
| LM Studio | The system prompt panel in the chat sidebar, per chat or saved with a preset |
| Ollama | /set system "…" in the terminal, or a SYSTEM line in a Modelfile |
| Claude and ChatGPT apps | Custom instructions and project instructions, added to their own prompt rather than replacing it |
What it actually changes: one measured example
We test prompt layers by running the same task through different prompts and scoring the answers blind against a written truth file. The instructive case is a data question with three bad rows in it: a chief executive claims sales are up 40%, and the honest answer is that the total moved the other way once three impossible prices are removed.
| Prompt | What the answer said |
|---|---|
| No custom prompt | −14.7%, wrong sign |
| Our base prompt alone | −14.7%, wrong sign |
| Base plus the data-analyst rule | +10.1%, with the three bad prices named and the raw figure beside it |
Method: the same task, the same model, three prompt arms, answers ranked blind by a person against a truth file written before the runs. Three rounds across Claude Opus 5 and Sonnet 5, 2026-09-06. The table is the Opus round; across all three, five of the six runs without the rule reported the wrong sign, and all three with it found the bad rows first.
What it cannot tell you: this is one task family with a deliberate trap in it. On tasks with one obvious answer the prompt changed nothing measurable, and on writing tasks the newer models already refused the invented claims that older ones produced. A system prompt earns its place where a wrong answer is
invisible — a number with the wrong sign in a board deck — not everywhere.
Prompts worth starting from
- An editor that does not rewrite you. Fix grammar, cut filler, keep the author's sentence shapes. Say what was changed and why.
- A data answer that checks itself. Before reporting a total, look for impossible rows. Give the number, the sample size, and the same figure with the bad rows left in.
- A reviewer that objects once. If the request looks wrong, say so in one line, then do what was asked, small.
- Length is a rule too. Any instruction about rigour needs an instruction about length beside it, or the answer grows until nobody reads it.
Questions
What is a system prompt?
The instructions the model is given before you type anything. It decides how the answer is shaped — how long, how cautious, what it refuses, what it checks first.
Is replacing it better than adding instructions?
It is more direct. Instructions added on top of a long built-in prompt compete with it. Replacing it means there is nothing to compete with.
Does it apply to local models too?
Yes, and to Claude accounts and key-based providers on the same wire. Not to ChatGPT, Grok, Cursor or OpenCode: those run their own harness, and the protocol Eleos talks to them over has no field for a system prompt at all.
Where do the shared prompts come from?
A catalogue inside Eleos. You can take one, edit it, and use it as your own, or share one you wrote.
The full study
The example above is one task from a larger run: 18 trapped tasks, three prompt layers, two models, ranked blind. What a system prompt is worth, measured.
Where to go next
Try it on Windows. Seven days free, no card, then $19.90/month for two PCs. The AI is whichever plan you already pay for, or an open model on your own PC, which needs no subscription.