Use /Goal for Long-Horizon, Self-Checking Execution

Some tasks should not be treated as a single response. They require several tools, intermediate artifacts, inspection, correction, and repeated work before the outcome is complete. The /goal command gives Hermes a standing outcome and allows it to continue across turns until the outcome is achieved or the execution budget is exhausted.

A goal should describe the completed result, not merely the first action. "Make a website" is an instruction. "Build and deploy a polished animated website that explains the offer, contains generated visuals, includes strong copy and calls to action, and returns a live link" is an outcome.

A goal creates a loop:

1. Interpret the desired outcome.
2. Decompose it into required tasks.
3. Use the available skills and connectors.
4. Create intermediate assets.
5. Inspect the result.
6. Identify missing or weak elements.
7. Revise the work.
8. Continue until the success conditions are met.

The website build shows why this matters. Hermes must define the offer, write copy, create a visual direction, generate images or video, assemble animated sections, test the page, correct problems, deploy the site, and return a link. A one-shot prompt can stop after an incomplete draft. A goal loop is intended to keep building, reviewing, fixing, and repeating.

The execution can also operate with a turn budget. In the website example, Hermes reported a 20-turn budget. This gives the goal room to continue while still creating an upper boundary.

Long-horizon work fails when the agent treats the first plausible output as completion. Websites, multi-asset marketing kits, edited videos, and cross-application workflows require quality control and dependency management.

The /goal pattern shifts the user's effort from supervising every step to defining the endpoint. The agent is responsible for the internal sequence, including review and repair.

This is especially useful when the user is working from a phone. The user can issue the outcome, allow Hermes to continue, and receive the deployed result or review artifact through the channel.

Write goals with five components.

1. Outcome. State what must exist at the end.
2. Quality criteria. Define the style, completeness, and user experience expected.
3. Required capabilities. Name the skills or connectors Hermes should use.
4. Delivery condition. Specify where the final artifact should be stored or deployed.
5. Completion evidence. Require a link, file, dashboard, or project status that proves the work is

finished.

For example, a website goal can require an animated background, polished motion, visual assets, offer copy, repeated booking calls to action, deployment through here.now, and a live link returned to the home channel.

A video-editing goal can require two passes: remove bad takes and pauses, then add B-roll, captions, and a final inserted clip. The completion evidence is the final exported video and a Multica issue moved to review.

A property-marketing goal can require a complete kit, a deployed page, a list of additional listings, and contact details collected through Firecrawl.

After a goal completes, inspect the artifact. If a defect remains, give a concrete correction and allow Hermes to run another improvement pass. When the result becomes repeatable, turn the goal procedure into a skill.

Goal loops consume more time and model usage than simple prompts. They can also continue in the wrong direction if the outcome is vague or the quality criteria are missing. A turn budget limits open-ended execution but may stop before a complex task is complete.

Self-checking does not guarantee perfect judgment. The generated website may still need copy changes. An edited video may retain black bars or rough transitions. The user remains responsible for reviewing the final result, especially before public release.

Use /goal when success requires multiple tools and revision cycles. Define the completed outcome, quality criteria, delivery location, and evidence of completion, then review the result rather than micromanaging every step.