Run your first task

Describe the task in a sentence, attach your material, and ClawsGO gets to work on a cloud host.

Write a good task description

No prompt engineering needed. Brief it the way you would brief a new colleague. Three parts:

Goal: what you want back (a review, a figure, runnable code).
Material: attach files or point to data sources.
Constraints: format, length, language, deadline.

Example

Using the 42 PDFs attached, write a systematic review on the stability of
perovskite solar cells. Around 8,000 words in English, organised into
degradation mechanisms / encapsulation strategies / testing standards.
Cite the source paper for every claim and finish with a comparison table.

The message body is capped at 32,000 characters. Anything longer goes in as an attachment.

Attach material

Three ways: "Upload files" in the "Add" menu of the composer, drag files onto the composer, or paste (screenshots included).

Up to 10 files per message. The per-file size limit depends on your plan, see Plans.
Uploads start as soon as you pick the files. You cannot send while an upload is in progress. Failed uploads offer "Retry" and "Remove".
Type @ to reference a file already in the task directory or an earlier task from your team (the agent reads its conversation as of send time; its files are not copied), or, on a cloud host, to name preinstalled software & tools and built-in databases (the plus menu lets you browse them by domain); type / for commands and skills.

What you can set before sending

The settings button in the composer shows the current model and reasoning level. It opens five groups:

Option	What it does
Mode	"Quick": direct answers, no tools, for simple questions. "Solo": one agent, end to end, for focused tasks. "Delegate": the main agent plans, subagents execute, for multi-step work; Plus and above. "Cluster": many subagents in parallel, built for large-scale surveys and deep research, with the highest credit usage; Pro and above. Switching modes inside an existing task asks for confirmation: the available tools change and the context cache resets once
Model	See Choose a model
Subagents	Shown in Delegate and Cluster modes. The model used by subagents and cluster members that the main agent dispatches. Defaults to the main model, see Choose a model
Reasoning	Low / Medium / High / Extra High / Max. Max uses significantly more credits per turn and may make the model overthink; if it runs too long, use High or Extra High
Context	256K, 512K or 1M, see Choose a model

All five settings are remembered: a new task starts with your last choices, and an existing task keeps the settings of its previous turn.

The Claude Fable line, Cluster mode, Max reasoning and 1M context use significantly more credits. Switching to them asks for confirmation and warns "High credit usage. Use with care." While on, they show in amber on the settings button. When a new task keeps them from last time, they are listed below the composer; "Reset to default" switches them back. In credit history, charges from turns run with Cluster, Max or 1M are marked.

Above the composer, pick a host, then either "Choose a project" or "Choose a folder" (one or the other; pick neither and a working directory is assigned automatically). Choosing a project opens its home page. Folders are only available on local devices. Once chosen, hover the icon to reveal an × that clears the choice. Tasks started inside a project use the project's host.

Enter sends, Shift+Enter inserts a line break. Once the agent has replied, the ring to the left of the send button shows how much of the context window the task uses; hover it for used and total tokens.

While the task runs

Every tool call appears on the timeline as it happens. While running, steps are expanded; once settled they collapse into a single line, "Worked through N steps", which you can open to see each step's input and output. Click a step that reads a file in the task directory to preview it on the right; click a step that reads a web page to open the page in a new tab. The model's thinking shows up there too, as a single "Deep thinking" line you can open to read it (a few models don't expose their thinking and only show the time spent). The context ring next to the send button updates as the agent works; earlier conversation is compacted automatically as it fills.

Keep working until a goal is met

Type /goal followed by a finish condition, for example /goal All tests pass and the report is saved as report.pdf. Each time ClawsGO is about to stop, a separate check reads the conversation and decides whether the condition is met. If not, it shows "Goal not met yet, continuing" with the reason and keeps working. Once met, the timeline shows "Goal met".

Write a condition the conversation can prove: name the file, the command output or the result you expect. The check reads only the conversation, not the files themselves. Asking for command output as proof makes the result more reliable.
Each check is one extra model call and uses credits.
After 9 checks in a row that fail, the reply pauses. The goal stays on the task: your next message continues toward it.
If a check times out or fails, the timeline shows "Goal check didn't finish" and the reply stops. The goal stays set, and your next message checks again.
Quick mode can't set a goal. Messages sent in Quick mode skip the check, and the goal stays set.
While a goal is set, it shows above the composer. Hover it to see the full condition and "Clear goal". /goal on its own shows the current goal; /goal clear clears it.

Where the deliverables are

Files ClawsGO delivers are listed under "Deliverables" in the Overview tab of the right panel, newest first; click one to preview it. Everything in the working directory is in the "Files" tab. When ClawsGO delivers a Markdown, PDF or image file it opens a preview automatically and the right panel folds away to make room; close the preview and the panel returns. See Files and previews.

After that

Send another message to add requirements. A task can keep going indefinitely.
If a task is interrupted, a reason card appears above the composer or at the end of the reply. Most of the time "Continue" resumes it. The full table is in Troubleshooting.
"More actions" in the header lets you pin, rename, share, hide or delete the task. From the sidebar you can also "Convert to project" so that follow-up tasks share one working directory.