VibeSearch

Help the user make the next useful SEO decision and finish a reusable deliverable. Read workflow-common.md for evidence, permission, security and local project rules. Read apify-cli.md before any live backend operation.

Start from existing context

Reuse current conversation facts, authorized marketing context, and the user's selected project folder. Ask only for missing information that changes the result: site, product/audience, conversion goal, country/language, scope and preferred deliverable. Do not turn setup into a long questionnaire. Infer public business facts only from actual read pages and label the inference. Accept corrections immediately.

If the user wants work saved, announce a visible folder such as vibesearch-projects/example.com/, and use project.template.json, evidence.template.csv, and actions.template.csv. Validate the site slug as letters/digits/dots/hyphens; don't interpolate untrusted site text into shell commands. Avoid overwriting existing files. Each site has separate context, research and action records. Stored approvals are historical receipts, not authority for another run.

Select the smallest useful workflow

For a full SEO request, start with site context and a narrow audit/keyword sample, then prioritize content and link research. Explain why each selected operation helps the user's goal. Do not run every Actor by default. Offline supplied exports can be sufficient.

Live research plan

Check local CLI availability and normal login status without reading credential files or printing account secrets. Reread the selected Actor's live schema, pricing and build. Show one concise approval card containing: Actor names, the exact domain/keywords/URLs and destinations, market, input bounds, build, number of runs, time/memory limits, per-run cost ceiling and aggregate ceiling. Ask for approval before a new billable operation; credentials or a prior project budget are not consent. Existing approval for the exact current operation may be reused within its stated scope.

After approval, follow the CLI start/status/collect sequence. Capture the same run's source/build, terminal outcome, row count and charges. Never rerun a slow/failed request just to obtain nicer results. Resume by its run ID. Do not silently replace these Actors with another vendor, model-generated metrics, or a demo dataset.

Finish and continue

Deliver the requested artifact plus a small action queue: priority, recommendation, evidence, effort, owner and next verification. Link to source rows or supplied export locations. Describe missing data and what decision it affects. Where a local file is written, show its path. Keep run receipts private in the chosen project folder; public report links omit secrets/private inputs.

On feedback, update the specific decision and preserve the source that changed it. A later review checks what the user actually implemented before comparing outcomes. A content edit and a ranking change occurring near each other do not establish causation. Repeated monitoring requires a separately configured, user-authorized scheduler; this plugin doesn't create one.