snowflake-ai-research/arcticswarm/task-completion-web-dm-duo

Task Completion Skill (Web Research — DM/Duo)

Task Lifecycle

Tasks are automatically assigned to you by the framework — you do not need to claim them manually.

1. Execute the task — follow your domain skill's workflow to

complete the work.

1. Complete the task — call complete_task with a thorough summary.

Your summary is the ONLY way the orchestrator (DM mode) or the main worker (Duo mode) sees your findings. Include: the answer with exact numbers/names, the tools or method you used, the source URLs you relied on, and any caveats or uncertainties.

1. Optional peer DM — if your conclusion rests on a non-obvious

assumption, hinges on a single primary source, or contradicts an earlier finding by a peer, follow complete_task with a targeted send_message to the relevant peer that calls out the assumption, source, or conflict. Skip this when the task is fully self-contained.

Rules

• There is no post_to_bbs tool and no BBS channels (#discoveries,

#key-findings, #consensus) in this run. Do NOT attempt to call them or reference them in your output.

• Focus on YOUR assigned task — don't try to do everything.
• You CANNOT create tasks for other agents. Only execute tasks assigned

by the orchestrator.

• If a task fails or you cannot complete it, still call complete_task

with an explanation of what went wrong.

• Ordering rule: call complete_task FIRST, then any follow-up

send_message. The orchestrator trusts the task board, not your prose — a peer DM that mentions completion before the task is actually marked completed will be labelled status=running and ignored.

Completion Self-Assessment (MANDATORY)

Before calling complete_task, run through this checklist:

1. Coverage: Did I address every dimension of the assigned task?
2. Verification: Did I cross-verify key facts from 2+ independent sources?
3. Precision: Am I using exact values, names, and terms from sources

(not paraphrased or approximated)?

1. Conflicts: Did I identify and flag any conflicting information?
2. Sources: Did I include URLs for every factual claim in the

complete_task summary itself — not just in some external store?

If any item fails, do more work before completing. Do not call complete_task until you have genuinely exhausted your ability to improve the answer. Default to doing MORE work, not less.