Prompt profiles

Nearly every sentence Mework adds to a conversation on its own, from the environment block to tool descriptions, tool results and notices, is a key with built-in English text. A prompt profile is a JSON file that overrides the keys you name; every other key keeps the built-in wording.

What a profile changes

A profile can reword or remove:

• the environment block, and the skills, MCP servers and hooks sections of the system prompt;
• the descriptions of built-in tools and of MCP tools;
• what tools report back when they succeed, such as memory and handoff confirmations and background task status;
• the notices Mework delivers mid-conversation, such as plan-mode guidance and the handoff request;
• the texts around memory, project instructions, subagents, workflows and web search.

A profile changes only what the model reads. It does not change:

• the conversation's own system prompt card. Edit that on the timeline, or in a preset's conversation template;
• which tools exist, their parameters (apart from a few parameter descriptions that have keys), or what needs approval;
• tool error messages, which are fixed English;
• markers Mework reads back, such as [Image #N] and <task-notification>, or the header line of MEMORY.md;
• the app's interface, including approval cards and dialogs, which stays in the app's language whichever profile you pick.

The built-in profile is Mework built-in. It is English, part of the app, updated with each version, and cannot be edited. Selecting no profile is the same as selecting it.

Where profile files live

Mework lists every .json file directly inside:

• ~/.mework/tool-descriptions/
• .mework/tool-descriptions/ in every local workspace you have added, not only the conversation's

Sub-folders and symbolic links are ignored, and workspaces on SSH machines or in WSL are not scanned. Mework never creates, changes or deletes these files.

Choose a profile

1. Open More options → Conversation settings → Advanced tools.
2. Pick the profile in Tool descriptions. Mework built-in comes first, then your files.

The same row is in a preset's settings, and new conversations copy the preset's choice (see presets). The list refreshes when you open conversation settings, or when you press Rescan on its Skills, MCP or Hooks page.

A selection is tied to where the file is. Changing name inside the file keeps it selected; renaming or moving the file, or its workspace folder, does not.

When changes apply

• The selected file is read again at the start of every run: edit it, save, and send your next message.
• New texts apply from the next run. Tool results and notices already in the conversation keep their old wording.
• Subagents and workflow steps use the profile the conversation had when they started.
• Switching profiles, or editing the selected file, mid-conversation changes the start of the request, so the next request cannot reuse the provider's prompt cache. After the conversation's first request, the Tool descriptions row is drawn orange like the other settings that cost the cache, and Mework asks once before switching.

File format

{
  "name": "Terse",
  "prompts": {
    "task.wait_idle": "Nothing is running.",
    "system.environment.date": "Date: {date}",
    "project_memory.untrusted_banner": ""
  },
  "tools": [
    {
      "toolName": "grep",
      "schemaNotes": "",
      "usageGuidance": "Prefer grep to reading whole files."
    }
  ]
}

If one tool has several rows, the first row with any text wins.

• Fallback. For each key, Mework uses the selected file's text when the file has the key, and the built-in English text otherwise. A selected file that is missing or not valid JSON means the built-in profile.
• Empty strings. "" means "say nothing here". An empty system.environment_section removes the whole environment block; an empty project_memory.untrusted_banner drops the banner from project instructions.
• Keys that ship empty. web.findings_notice, web.results_notice (a notice on web search and fetch results) and web.untrusted_marker (a prefix for retrieved lines that look like instructions) add nothing until you fill them in.
• Placeholders. Words in braces, such as {date}, are filled in when the text is used; the key reference lists each key's placeholders. You can leave one out. A placeholder the key does not declare stays in the text as written.
• Lists. When Mework joins names into a list, it puts format.list_separator between them (, by default).
• Size. The file must be UTF-8 JSON, at most 64 KiB.

Write your own

Start from the built-in profile, which has every key with its current text, and the key manifest, which lists each key's placeholders and where it appears:

• prompt-profile.en-US.json — The built-in profile
• prompt-profile-keys.json — The key manifest

1. Download prompt-profile.en-US.json and save a copy as ~/.mework/tool-descriptions/<name>.json, or in a workspace's .mework/tool-descriptions/.
2. Set name, then delete every key you are not changing. Keys you leave out keep following Mework's updates; a key you keep stays at the text you wrote.
3. Edit the texts you kept, keeping the placeholders each one uses.
4. Open More options → Conversation settings → Advanced tools and pick the file in Tool descriptions. If it is not listed, close conversation settings and open them again.
5. Send a message, then open More options → History. The system rows show the system prompt Mework sent, and the tools row shows the tool descriptions.

To have a model work in Chinese, or any other language, translate every key and keep the whole file: a key missing from the file falls back to English.

Key reference

Every key, grouped by its prefix, with the placeholders it accepts and where its text appears. Open a group to see its keys.