Preface¶

The core philosophy of DeepSeek Harness (dsh) is “Everything is a plugin”: models, tools, skills, and conversations can all be replaced by community packages. Agents excel at presenting complete plans in one go, but for those who struggle with execution, the plan itself often becomes a burden—the longer the checklist, the harder it is to get started.

dsh-adhd-copilot was built exactly for this scenario. It does not require you to first practice self-discipline and then complete everything all at once. Instead, it scans for states like “procrastinating / overwhelmed / self-blaming”, routes to the corresponding module, and only provides 1 to 3 actionable minimal steps each time. This article is compiled after cross-checking the community directory page, the GitHub repository README, SKILL.md, and the source code: explaining what it is, how to install it, and how to use it in practice.

The community plugin directory https://deepseek-harness-plugin.com/ is an independent site and has no official affiliation with DeepSeek or Fangxin. Do not treat it as an official app store.

What it is¶

dsh-adhd-copilot is a skill plugin for DeepSeek Harness, maintained by zimai233, with its repository at GitHub: zimai233/dsh-adhd-copilot. The license is MIT, the package.json version number is 0.1.0, and the primary language is JavaScript. As of 2026-08-18, both the directory page and GitHub show 3 stars.

The positioning on the directory page is: an ADHD behavior coach skill that uses a gentle, actionable structure to help you advance tasks. The repository README is more specific: it takes users from “wanting to start but being unable to move” to “taking away one minimal action”.

It does not solve the problem of “helping you write a weekly report outline”, but rather these stuck moments:
- Unable to start even though you want to (activation difficulty)
- Too many things to do, not knowing where to start first (overwhelm)
- Easily distracted, unable to sit still
- Underestimating task duration
- Unmotivated, bored, lacking energy
- Self-blaming for not completing tasks
- Not knowing how to start a task

At runtime, it only registers one tool for the agent: adhd_breakdown. The source code comments note: referring to the unconfirmed skills service in the agent, so the plugin only registers tools for more stability; the complete coaching methodology is still included in the files field of the npm package as independently readable deliverables, exactly as it appears in skills/adhd-copilot/SKILL.md and 11 references/*.md files.

Core Features¶

Three Principles¶

SKILL.md and the README share three driving philosophies:
1. Always only push for the next step, without requiring full completion.
2. Actionable takeaways: Each reply provides a minimal action that the user can perform independently, no need to return to the chat window; feedback loops are entirely optional.
3. Quality floor: Content quality takes precedence over formatting; when rules conflict with good answers, the rules give way.

references/core-workflow.md also lists five cognitive facts explaining why steps must be short and time estimates must be broken down into minutes: working memory is limited, knowing does not equal doing, activation is the hardest part, time perception is distorted, and visible progress matters.

One Tool, Eight Modules¶

When the agent calls adhd_breakdown, it scans for state signals in task and context, and routes to modules in the priority order of **emotion > structure > activation`. The order in the source code is: first handle self-blame (F) and low motivation (E), then clean up overwhelm (B), and finally handle activation (A).

M5 “Interest Gamification” does not follow the default routing, and will only be packaged into RPG / serious / minimalist narrative when the user explicitly opts in.

The structured fields returned by the tool are clearly explained in the README: module / moduleName, nextAction (can be done within 30 seconds), whyNow, timeboxEstimate (specific minutes, no “a while”), energyRequired, steps (1-3 steps). Optional parameters include: battery (M3 energy level: 10 / 30 / 60 / 90), narrative (M5), estimatedMinutes (M6), parked / dropped (M2’s later / abandoned list).

Pre-flight Gate and Output Discipline¶

Not every message will go through the coaching process:
- No task signals (small talk) → silent, zero interference
- Small request (one-sentence question) → direct answer, no process wrapping
- Ambiguous signals → provide an A-G self-help menu for the user to choose by themselves, reducing misjudgment
- Clear signals → directly enter the corresponding module

The output discipline is written in references/output-discipline.md: the first line is the next action; numbered steps do not exceed 3 at a time; end with one more action that can be completed within 30 seconds; praise must be tied to specific wins, and phrases like “come on, you can do it” are forbidden; AI cannot actively appear during execution, so never fake a timer.

It is implemented as pure ESM, with zero build steps and zero runtime dependencies. package.json declares the Cordis layer via "dsh": { "bundle": { "patch": "./cordis.patch.yml" } }, and the patch file inserts the plugin with id: adhd-copilot into the current profile.

Installation and Activation¶

The installation command given on the community directory page can be run in a terminal where the dsh CLI is already installed:

dsh plugin add github:zimai233/dsh-adhd-copilot

The official CLI habitually includes the profile. The写法 from the repository README is:

dsh plugin --profile myprofile add github:zimai233/dsh-adhd-copilot

For reproducible installations, pin the commit hash. The format on the directory page is github:zimai233/dsh-adhd-copilot#commit. As of 2026-08-18, the latest commit on the repository’s master branch is 6527cb64b11b7265ba576e195bd008a5aa066be5:

dsh plugin add github:zimai233/dsh-adhd-copilot#6527cb64b11b7265ba576e195bd008a5aa066be5

The README also mentions dsh plugin --profile myprofile add dsh-adhd-copilot (installing via npm package name). When checking the npm registry, dsh-adhd-copilot returns a 404, so you should use the GitHub source as the current standard, and do not default to npm.

Both the directory page and the dsh documentation remind users that the plugin runs with the permissions of the current dsh process, and may execute code during installation. Please inspect the source code repository and license before installing.

Typical Usage¶

After installation, you can speak naturally to the agent without manually writing tool calls. The README provides three trigger phrases:

“I’ve been procrastinating on writing the weekly report for two weeks, help me get started.”

“So many things piled up, my mind is a mess, not knowing where to start first.”

“I didn’t finish it again, I’m so terrible.”

The first phrase will hit M1. The structured breakdown example in the README is:

{
  "module": "M1",
  "moduleName": "任务破碎机 (Task Breaker)",
  "nextAction": "Open the file or page related to "weekly report" (within 30 seconds)",
  "whyNow": "Activation is the hardest part: the first action must be obvious, tiny, and doable right now. Only give 1-3 steps at a time, each under 60 seconds, and ask for the next step after completing this round.",
  "timeboxEstimate": "First step <60 seconds; 2 steps in this round, ask for next step after completion",
  "energyRequired": "Low (only need to open a file)",
  "steps": ["Open the file or page related to "weekly report" (within 30 seconds)", "Write this line as the title: "Weekly Report""]
}

After being rendered into a coaching reply, it roughly looks like:

Weekly report, don’t think about finishing it yet. Just do the first step:
1. Open the file or page related to “weekly report” (within 30 seconds)
2. Write this line as the title: “Weekly Report”

Each step takes less than 60 seconds. Reply “Done” after finishing (optional, not required to proceed). You’ve started, and this step is already a win.

The third phrase will hit M7 (Failure Restart). The corresponding reply in SKILL.md is: first stop self-judgment, then provide one 30-second action—just open the relevant file, without requiring the task to be fully completed.

The repository includes npm test (node --test), which covers state scanning, Pre-flight silence, ambiguous menus, and whether the skill files are complete. If you need to verify the breakdown logic, you can run the tests directly without first connecting to a real conversation.

Applicable Scenarios and Notes¶

It is suitable for users who are already using DeepSeek Harness and often get stuck on “activation / overwhelm / self-blame” during execution. It is not a to-do app, and will not supervise your Pomodoro sessions for you. The “Honest by design” section of the repository states plainly:
- No timer: AI cannot appear during execution, and mid-process checkpoints are all optional.
- No guaranteed cross-session memory: Persistence only occurs when the platform itself provides long-term memory.
- Zero interruption: Completely silent when there are no task signals.
- Safety boundary: This is an execution aid, not a medical diagnosis or treatment tool. If the user shows signs of emotional crisis or self-harm, gently suggest seeking professional help instead of continuing to push tasks.

SKILL.md also lists several anti-patterns: after installation, if the agent still throws a complete checklist at once, gives twenty options for you to choose from, or says “you should be more self-disciplined”, it means it is not following this discipline.

The plugin runs with the permissions of the current dsh process. Before installing, please visit https://github.com/zimai233/dsh-adhd-copilot to view the source code and MIT license; for a reproducible environment, pin the commit, do not just install the floating default branch.

Closing¶

dsh-adhd-copilot packages a complete ADHD coaching workflow into a single tool in DeepSeek Harness: first identify the state, then only push the next step. The complete methodology is in skills/adhd-copilot/, which can be read as documentation or loaded along with the plugin.

Directory page: https://deepseek-harness-plugin.com/zh-CN/plugins/dsh-adhd-copilot/

GitHub: https://github.com/zimai233/dsh-adhd-copilot