CLAWPOD ENGINEERING

Choose a Chat Mode: Plan First or Carry Out Agreed Work

Separate Default, Discuss, and Execute, understand when a selection applies, and turn a classification plan into a result you can verify.

ClawPod Engineering · 6min read
REVIEWED BYEditorial Review · Capability Review
A desk with an open planning notebook and neatly arranged documents suggesting execution
An AI-generated metaphor for planning and organized work, not a product screenshot.

You do not want an Agent to reorganize files before you have agreed on a classification rule. Once the rule is settled, you also do not want an endless comparison of alternatives. This guide helps people working in a ClawPod Cloud Chat Room choose Default, Discuss, or Execute and turn a plan into verifiable work. A Room is a shared conversation context for users and Agents; you need access to the Room and the materials involved.

The distinction is practical: Default can make changes, Discuss focuses on research and planning, and Execute focuses on carrying out agreed work and verifying results. In every mode, put the target, prohibited actions, and completion criteria in your message.

A mode changes the approach, not your permissions

The current official Guide describes the modes as follows. This is not an experimental comparison of response quality.

Mode Purpose described by the Guide When to choose it
Default Handle questions, summaries, and requested changes The task is simple and clear
Discuss Research evidence, organize requirements and alternatives, and make a plan You have not decided what to do or how
Execute Carry out agreed work systematically and verify results You want to turn a plan into work and checks

Default can make actual changes. Do not treat it as read-only. Changes in Discuss are limited to outputs used for planning, such as a planning document. To implement the plan, switch to Default or Execute and send a new request. A simple summary does not need a mandatory discussion-then-execution sequence.

Selecting a mode or approval setting does not add tool permissions, expand organizational access, or approve unrequested external actions. Execute does not guarantee completion when required tools or permissions are missing or external approval is pending.

Your selection applies to the next message

The product surface in this article is the Chat Room composer in the ClawPod Cloud Portal. This is not a guide to settings in the Agent Control UI. The steps below are documented by the official Guide; they are not confirmation of the live screen or menu availability in every deployment.

  1. If the composer shows the chat mode menu, choose Default, Discuss, or Execute. The Guide also describes switching with Shift+Tab in the composer.
  2. Open customization only when you need a specific setting. Leave it on Auto if you are unsure.
  3. Write the materials, desired output, prohibited actions, and completion criteria, then send the message.

Changing the selection alone does not start work. The selected mode and settings apply when you send the next message. Do not treat the selector as a control that stops or restarts work already underway.

According to the Guide, settings are stored in the current browser per user and Room. Check them before sending if you move to another browser or Room. If the menu is absent, follow the ordinary message-writing procedure. Absence alone does not establish a configuration error or justify guessing hidden URLs.

Illustrative example: agree on categories, then produce a list

This is an illustrative example, not an observed run or a guarantee of automatic success. Assume approved material A is an installation guide, B covers incident response, and C is an installation checklist. The task is to create a classification list in the Room without moving the original files.

Discuss request

Read the supplied A, B, and C and propose classification rules and a plan. Do not modify, move, or delete originals or send them externally. Ask rather than guess if the content is unclear.

Illustrative planning result

I propose Installation and Incident response as categories. A and C belong to Installation; B belongs to Incident response. I will check that all three materials appear and that none appears twice, then present the classification list in this Room. I will not change the originals.

After reviewing the rule, choose Execute and send a new message.

Create the classification list for A, B, and C using the Installation and Incident response categories we agreed on. Include each classification reason and checks for omissions and duplicates in this Room. Do not change originals or send materials externally.

Illustrative execution result

Material Category Reason
A Installation Describes the installation sequence.
B Incident response Describes incident symptoms and response steps.
C Installation Describes post-installation checks.

I compared three inputs with three results. There are zero omissions and zero duplicates. I did not change originals or send materials externally.

Judge success by comparing the materials and output, not by the confidence of the closing sentence. Check that A, B, and C each appear once and that each reason matches the actual content. Where necessary, verify unchanged originals separately through file comparison or work records. The sample sentence “I did not change originals” is not proof of that postcondition.

If B cannot be read, do not accept “all complete.” Ask for a report separating the inspected A and C from unverified B, provide accessible material, and check the missing portion again. If no result has arrived, check current progress and remaining conditions before repeatedly switching modes or sending the same work again.

Customize only what you need

The Guide lists Auto, Lite, Deep, and Full for discussion depth. Auto chooses depth for the conversation. Lite asks only essential questions and plans quickly. Deep asks about key decisions and compares alternatives. Full explores broadly, one question at a time. Do not interpret these names as a ranking of quality or accuracy.

For Execute, keep these settings distinct:

  • Task management tool: Auto chooses an appropriate recording approach; a separate work record may be unnecessary. Tasks manages work, assignees, and progress in a shared project. Workboard manages cards and execution state on the Agent’s work board. Choosing a management tool is different from specifying sequential or parallel execution.
  • Review requirement: Auto determines whether independent review is needed based on the work and its risk. Required specifies review by an independent Agent. If review cannot be performed, do not count it as completed; check the reason and remaining conditions.
  • Progress check: Auto or 5, 10, or 30 minutes requests a follow-up check on waiting work. It is not a schedule that repeats every task at that interval. Verify actual follow-up results in the Agent’s report.
  • Approval approach: Auto follows user instructions and applicable approval rules. Within approved scope proceeds within explicitly approved work and impact, which can include approved risky actions. Confirm on risk asks again immediately before risky actions, even within approved scope.

Read the result for what was actually recorded, reviewed, and checked rather than relying on the selected setting alone.

Keep additional instructions reusable

The Guide describes up to 240 characters of additional instructions per mode, sent with each message in that mode. A repeatable instruction such as “Write the result in English and separate changed items from verification results” belongs here.

Put one-off scope such as “Process only A, B, and C” in the message body. Do not put passwords or API keys in additional instructions. A short additional instruction is not a reason to omit targets, prohibited actions, or completion criteria from the message.

Before your next request, decide one thing: Do you need a plan, or the result of agreed work? Choose the mode accordingly and include what you will compare the result against.

Sources