Preface¶

When building writing agents in DeepSeek Harness (DSH), a common practice is to put “please imitate my style” into prompts, rules, or session instructions. After switching models, tools, or starting new sessions, these style requirements tend to become scattered and need to be rewritten repeatedly.

Liltloom (语织) is a Chinese-first, user-controlled AI writing style memory layer. It silently learns language preferences from the user’s own expressions that pass qualification checks, saving them as structured, reviewable style rules; by default, it is explicitly activated, and only generates a bounded Style Context Packet after being enabled.

What This Is¶

Liltloom is maintained by Adkid-Zephyr and licensed under MIT.

It extracts writing style from temporary prompts and turns it into checkable, editable, and portable rule data. Currently, the officially maintained adapter is the DeepSeek Harness Web adapter; in 0.2.0, only the DSH adapter has reached an installable, testable state.

It mainly solves three types of problems:

• Style requirements are scattered across multiple prompts, sessions, or tool configurations;
• Users find it difficult to view, modify, or delete what the model has “learned”;
• Users want to export and migrate style rules rather than being tied to a specific session.

Core Features¶

The following are verified plugin capabilities:

• Chinese-first, user-owned writing style memory layer;
• Silently learns language preferences from the user’s own expressions that pass qualification checks;
• Saves structured, reviewable style rules;
• By default, explicitly activated; generates a bounded Style Context Packet only after being enabled;
• Rules can be viewed, added, edited, locked, ignored, and deleted;
• Supports readable JSON export, import, and migration;
• Provides DeepSeek Harness Web adapter and native settings interface;
• Provides /liltloom command, trusted plugin interface, optional model tool, and loopback RPC;
• Optional Deep Style original text snippets and small model analysis, both disabled by default or require authorization.

Basic mode does not save full conversation transcripts by default. Deep Style is disabled by default; when enabled, it retains sparse original text snippets subject to quota limits. Small model analysis requires explicit configuration of a separate provider/model route and first-time user authorization.

The model tool is not registered by default; it only appears after executing /liltloom tool on.

Installation and Activation¶

The following environment requirements must be met before installation:

• Node ^22.19.0 || >=24;
• pnpm;
• A DSH Web profile providing storage-domain.

The installation commands given in the README main workflow are:

git clone https://github.com/Adkid-Zephyr/Liltloom.git
cd Liltloom
pnpm install
pnpm run check
dsh plugin --profile web add "$PWD"
dsh --profile web --dump-config
dsh --profile web

Among these, dsh plugin --profile web add "$PWD" registers the current directory to the DSH Web profile; dsh --profile web starts the Web profile. Commands for source code execution and tarball installation are also described in the README.

Typical Usage¶

After starting, open:

Settings → Plugins → Liltloom

Here you can manage:

• Overview;
• Rules;
• Preview;
• Advanced;
• Data.

Commonly used /liltloom commands are as follows:

/liltloom preview
/liltloom use session
/liltloom off
/liltloom mode explicit|auto|disabled
/liltloom deep on
/liltloom export

The purpose of each command:

1. /liltloom preview: View the complete Style Context Packet that will be used.
2. /liltloom use session: Continuously use the style in the current session until /liltloom off is executed.
3. /liltloom off: Disable current style activation.
4. /liltloom mode explicit|auto|disabled: Switch activation modes.
5. /liltloom deep on: Enable the original text snippet capability.
6. /liltloom export: Export readable JSON.

Data Migration and Deletion¶

Liltloom supports readable JSON export and import, suitable for manual editing, backup, and migration.

To completely clear personal data, first execute:

/liltloom clear confirm

Then execute:

dsh plugin --profile web remove liltloom

Note: Uninstalling the plugin does not automatically delete personal data; explicit confirmation is required to delete personal data.

Applicable Scenarios and Considerations¶

Suitable for the following usage patterns:

• Using the DSH Web profile and wanting to maintain Chinese writing style rules;
• Wanting style rules to be reviewable, editable, lockable, deletable, and portable;
• Wanting to explicitly activate style only when needed, rather than injecting style information into every writing task by default;
• Wanting to add a user-controlled style memory layer within the DSH plugin ecosystem.

Things to note:

• The plugin runs with the permissions of the current dsh process; review the source code and license before installation;
• The currently officially maintained adapter is the DeepSeek Harness Web adapter;
• In 0.2.0, only the DSH adapter has reached an installable, testable state;
• Deep Style and small model analysis both involve more sensitive data or external calls, and are disabled by default or require authorization;
• Uninstallation does not automatically delete personal data; execute /liltloom clear confirm before deletion.

Links¶

• GitHub: https://github.com/Adkid-Zephyr/Liltloom
• Directory page lead (source not fully verified; may be regarded as a community directory lead rather than an official app store): https://www.skillhub.cn/plugins/Adkid-Zephyr/Liltloom