Search the work your AI agents already did

Try: Install Belt in the coding agent you already use. Add one recurring workflow as a skill, then see whether Belt suggests it when a matching task appears in a later session.

Try: Limit the first import to one real project, then search for a decision or feature name you abandoned. Read the surrounding messages to recover why it was rejected before an agent proposes it again. Keep the optional cloud, AI-insight, and sharing features off if those sessions contain private work.

Try: Use mq in read-only mode on the longest handbook or research file your agent sees regularly. Extract one named section, compare it with the source, then use the smaller result for a real task. Review any rewritten Markdown before saving because its update mode can change escaping.

Try: Run graft build and graft map on a copy of a real repository, then ask one question whose answer you already know. Check the returned files before trusting the map. Avoid graft init for now because open issues show that its generated ignore rule and agent configuration can be wrong.

Try: Start with one non-critical domain and one exact address rather than a catch-all. Forward it to a labelled Gmail address, then give your agent read and draft access through the CLI or MCP server. Keep approval and sending separate until the route, thread history, and From address all survive a real reply.

Try: Pick a year that matters to your product or audience and ask for two competing directions from that year’s research. Make the agent name the historical references and anti-clichés behind each one before it designs anything. Choose one direction, then apply it to a real page, ad, or social card at its final size.

Try: Run it after your normal tests pass. Make the agent trace callers and state changes, then reproduce every important finding before you keep it. Drop style comments and any warning that cannot name the exact input or timing that breaks the code.

Try: Give the skill one decision you are genuinely willing to change. Name the smart option you rejected and the downside you accept. If the opposite sounds ridiculous or your choice has no cost, you have an aspiration rather than a decision.

Try: Follow the five-step loop on one real search opportunity: find, inspect, propose, verify, then measure. Keep crawl, Search Console, analytics, and live-result evidence separate so each claim can be checked.