prakhar1114/ai_mime_marketplace/search-on-x

Search On X

Use this skill when the user wants to search X/Twitter for an exact phrase or term with date filters and extract structured post data.

Preconditions

• Chrome must be able to access X.
• If X requires authentication, the user must already be logged in to X in Chrome.

Inputs

The skill reads a JSON object:

• query: required exact search term, for example computer use.
• since: optional lower date/datetime bound. Defaults to one day before until or today. Date format: YYYY-MM-DD; ISO datetime is accepted.
• until: optional upper date/datetime bound. Defaults to today. Date format: YYYY-MM-DD; ISO datetime is accepted.
• limit: optional maximum post count. Defaults to 100; supports values up to 1000. Small values like 10 are useful for validation.

Run

Run with:

./run.sh inputs/inputs.example.json

run.sh uses the first available interpreter in this order: skill .venv/bin/python, workflow .venv/bin/python, then required $AI_MIME_PYTHON_PATH.

No requirements.txt is needed for this skill. If one is added later for build or manual repair, install it with:

"$AI_MIME_UV_PATH" venv .venv --python "$AI_MIME_PYTHON_PATH"
"$AI_MIME_UV_PATH" pip install -r requirements.txt --python .venv/bin/python

Those install commands are for skill build or manual repair. Runtime does not create or repair .venv.

Outputs

The script prints a JSON object to stdout with:

• query, since, until, limit
• search_url
• count
• posts: list sorted by likes descending

Each post contains:

• name
• username
• likes
• comments
• views
• url
• content

Content is the visible X search timeline content. Very long posts may be truncated by X with a visible "Show more" affordance.

Progress log format

Progress is emitted to stderr as JSON lines:

{"event":"step_start","id":"run_search","title":"Search X and extract posts"}
{"event":"step_done","id":"run_search","outputs":{"count":10,"search_url":"..."},"summary":"Extracted 10 posts"}
{"event":"workflow_done","outputs":{"count":10,"search_url":"..."}}

The browser harness also emits human-readable scroll progress to stderr, for example scroll 4: collected 13/50.

When fewer results are available than limit, the runner stops early in this order:

• immediately when X shows its no-results empty state;
• when the main search timeline is exhausted, with no loader and blank space after the last visible post;
• as a last fallback, after 3 consecutive scroll cycles add no new posts.

Fallback

If the script fails, read references/fallback_plan.md. The fallback is to open the generated X search URL or X Advanced Search manually, use the Top tab, scroll the timeline, stop when X shows no results or the timeline is visibly exhausted, and extract fields from visible posts.

ask_llm decision points

None. The skill uses deterministic browser-harness DOM extraction and numeric sorting.

References

• references/fallback_plan.md: manual/UI-agent fallback steps, selectors, URL format, and extraction notes.