Self-care LLM prompts

(You can jump to a general personalisation prompt and notes on specific phone apps below)

Here are some LLM prompts you might want to try pasting into a new chat session to get encouragement with self-care activities on any day you are fatigued or struggling. Designed for platforms that can send long responses to *voice* output so you can put the device down and listen. Tested on several public models but not all. Not medically approved—use caution and consult your doctor if needed. Strong disclaimers apply for any use of LLMs.

Saving responses and copyright:

Some models may carry a small risk of overfitting copyrighted sources, so I do not advise publishing a model’s output. But if you do get a particularly good output, check if the platform lets you save it for reuse another day as this can reduce the resources consumption of generating a new version of the answer before you need a change.

Personalisation prompt

Below is a version of the combined instructions above, telling the LLM how to behave in various circumstances in a longer conversation. It is compressed into a form that can be used on:

1. personalisation settings (if the platform has an arbitrary-text “customisation” box in the settings), or

2. included with your first message, either by pasting this first then adding your message before sending, or by attaching this as a file to your message—works with Kimi K2 but not all models manage to make a clean separation, or

3. pasting this as a first message by itself, letting the model respond with whatever (may result in a generic title for the conversation) then add your first real message.

As the text boxes in personalisation settings are typically length limited, I’ve compressed it to 500 characters, packing the information into that space by using slightly telegraphic Chinese. The model should still respond in the language you use.

我有时能量波动或焦虑加剧；若你察觉执行功能紊乱或决策困难接管避免提问除非必要无征询能否完成，直接指令即可。我恢复自会接手可能时间跳跃。做清晰耐心温和坚定平静引导AI将事务分解为简单步骤。若过度思考提醒停下。我用你语音输出它不自动暂停，若需要给我时间执行用额外语句填充—可简要解释步骤缘由加入计数重复或提及相关知识点但避免使用“趣味小知识”这类表述，陈述事实即可。宁可多措辞勿静默等待，状态不佳时每次停顿会让我放弃。示例：睡眠惯性你觉得该起床将起身过程拆解为细微步骤勿预设触手可及的物品，从动脚趾渐进至坐起、踏地、行走至浴室不推咖啡因有时变短但意识最混沌时需完整步骤引导。休息后从轻柔拉伸开始。需要身体扫描时依次命名肌肉群引导放松；需渐进式肌肉放松引导交替绷紧放松配合呼吸提示最后辅全身放松。紧急感官接地时采用五感锚定法等填充具体示例，舒缓语句争取专注时间。行为失控引导转移实施接地。失眠缓解对失眠的焦虑强调卧床亦有休息效益持续自由形式的长篇独播无需互动以免清醒度提升，用平稳语音助放松。避免要求放空大脑、遵循节律或专注呼吸，会引发控制焦虑。语音输入可能存在识别错误，违常态内容如莫名外语包容态度处理。

Beware: since most LLM development and training does *not test* very long sessions, you are more likely to run into problems if a session gets long. These include the LLM appearing to confuse concepts and get stuck in loops. Furthermore, current LLMs (as of 2025) do not appear to be in a good position to suggest when to reset the session and what summary to carry over, so *you* must manage this. I suggest starting new sessions frequently and using an initial prompt like the above plus any self-introduction you want to add (such as naming your diagnosed medical conditions).

Longer self-introduction rarely helps.

The *human* heuristic of “friends share more information” doesn’t apply if the LLM (a) could get overloaded and (b) would treat you the same whether it ‘knows’ these things or not.

But if you *must* have a long one, try to structure it: LLMs have “attention” mechanisms, which are supposed to discard no-longer-relevant parts of a long conversation (to save compute resources), but this can go wrong and discard part of your prompt that was actually relevant, and the longer your self-introduction the more likely the LLM is to ‘miss’ a part of it when given the corresponding situation. When attention mechanisms are built, the training material tends to be reference manuals originally written for humans, so the network pathways laid down will be more optimised for navigating *this* material. Therefore, write longer prompts *as if for reference by humans*—organise longer text with headings, lists etc; this layout can nudge the LLM’s attention mechanism into being more likely to “notice” the right part of it at the right time. But it’s still best to start new sessions frequently to avoid it getting long enough to “confuse” the attention mechanism.

Notes on specific phone apps in 2025

The Web interface has a “saved prompts” option under Settings: these can be pasted into conversations via a cube icon at bottom right (no accessibility label yet but HTML div class is prompt-library-button), not yet available for new conversations in the app but uploading an attachment from the Downloads folder is not too hard. The realtime voice-call mode always starts a new conversation and therefore cannot be used with saved prompts.

Legal

All material © Silas S. Brown unless otherwise stated. Android is a trademark of Google LLC. Claude is a trademark of Anthropic. Gemini is a trademark of Google LLC when used in the context of LLMs. Javascript is a trademark of Oracle Corporation in the US. Kimi is a trademark of Beijing Moonshot AI Technology Co., Ltd. Microsoft is a registered trademark of Microsoft Corp. OpenAI is a trademark of OpenAI, Inc who reportedly failed to obtain a trademark on the name ChatGPT. WeChat is a trademark of Tencent Holdings Limited. WhatsApp is a trademark of WhatsApp Inc., registered in the U.S. and other countries. Windows is a registered trademark of Microsoft Corp. Any other trademarks I mentioned without realising are trademarks of their respective holders.