Introduction

When an agent encounters uncertainty during execution, it immediately asks questions, fragmenting the user’s day. A dozen or so fragmented questions can ruin an entire morning. This is not just interruption; it also masks a pattern: most questions come from the same decision context. The ask-batch skill fixes this problem with a workflow of storing questions first, grouping them, and asking for answers all at once.

What It Is

ask-batch is a skill for DeepSeek Harness (DSH), maintained by ChenneyZhuang (MIT licensed). Its core is to collect open questions into a running disabled list, group them by decision context, and ask for answers in a single prompt that includes proposed defaults and an explicit answer format.

Core Features

  1. Store first, do not ask: Questions are not displayed immediately. Instead, they are added to the running disabled list (usually in session notes or a questions.md file), with a context line for each question. Work continues.
  2. Group by decision context: Questions are categorized. For example, all questions about the “report title field” are grouped together, while “database questions” and “lunch arrangements” are placed in different groups.
  3. Propose defaults: Each question or group includes a default answer, such as “Unless you say otherwise, I will assume X.”
  4. Ask once per group: Send one message per group, containing numbered questions, a designated answer format (such as “1: A/B, 2: free text”), and a deadline after which the defaults apply.
  5. Distinguish blockers from uncertainty: Only true blockers (questions that cannot proceed without an answer) are asked immediately and separately; other uncertain questions wait for batching.

Installation and Enablement

In a DSH environment, install it using the following command:

npx skills add ChenneyZhuang/ask-batch

After installation, the plugin runs as part of the DSH process. For agent-specific compatibility paths, see COMPATIBILITY.md.

Typical Usage

During a task run, the agent encounters multiple uncertain points:
* The agent stores these points in the disabled list and continues working on the unblocked portions.
* The agent analyzes the context and finds that “color scheme” and “Logo size” belong to the same decision context.
* The agent sends a batch message listing these two questions with default values and asks the user to confirm within one hour.
* The user confirms the defaults in one round of conversation, and work continues.
* If a blocker appears midway that requires an immediate answer, the agent raises it separately; the remaining questions stay waiting for batching.

Applicable Scenarios and Considerations

It is suitable for users who need deep work and want to avoid frequent context switching. The plugin runs with the current DSH process’s permissions. Before installation, review the source code and license.

One point to note:
* Over-batching is a failure mode: If a stale question is blocking a line of code, it should be asked immediately.
* Defaults cannot replace reading the room: In emotionally charged situations, defaults cannot replace direct asking or earlier interaction.

Summary

By aggregating fragmented questions into a single decision, ask-batch allows users to complete work with one response rather than being interrupted and fed one question at a time. For more details, see the plugin directory page or source repository.