x-search

x-search: searching X (Twitter) through xAI

This skill only matters when an xAI account is connected. The x_search tool

is registered at load time if a credential exists, and it is hidden behind

tool_search until you activate it. Everything below assumes you've read the

current date from the system prompt; never hardcode a date.

Which tool

1. Run tool_search "X posts". If x_search shows up, it's callable on the

next turn.

1. If nothing matches, xAI isn't connected. Say so plainly ("x_search is

unavailable, no xAI credential") and fall back to web_search. Don't guess

at tweets from memory.

x_search is one round trip to xAI's server-side X tools (keyword search,

semantic search, user search, thread fetch). You don't pick the sub-tool; the

server does. Your job is to shape the query and the parameters so the server

lands on the right one.

The 4 rules

1. Date-bound anything time-sensitive. Stocks, earnings, market reaction,

foreign flows, breaking news, product launches: compute yesterday from

today's date and pass it as from_date. Start narrow. If the first pass

comes back thin, widen to 7 days. Dates are inclusive on both ends.

1. Scope trusted accounts. When the user names accounts they trust, pass

them as allowed_x_handles (bare handles, no @, max 20). Example:

["growth_papa", "aleabitoreddit"]. allowed_x_handles and

excluded_x_handles are exclusive; pick one.

1. Phrase for recency. Write the query the way X advanced search expects:

"latest" or "recent" wording plus operators (since:, until:, from:,

filter:links, -filter:replies, lang:). The server has been observed

running keyword search in Latest mode when the query carries since:.

1. Split, don't stack. One mega-query returns mush. Run 2 to 3 narrower

searches: one pass by account, one by keyword, and a thread fetch when a

single post needs its replies. Each call is one search turn.

Operator cheat sheet

| Operator | Meaning | Example |

|----------|---------|---------|

| since:YYYY-MM-DD | posts on or after the date | since:2026-09-02 |

| until:YYYY-MM-DD | posts on or before the date | until:2026-09-03 |

| from:handle | posts by one account | from:growth_papa |

| filter:links | only posts with links | earnings filter:links |

| -filter:replies | drop replies | NVDA -filter:replies |

| lang:xx | language code | lang:ko, lang:en |

| "exact phrase" | literal match | "foreign net buying" |

Operators live inside query. from_date / to_date / allowed_x_handles

are separate parameters and are the reliable half; operators nudge the server,

parameters bind it. Use both when they agree.

Split-search recipe

Run these as separate x_search calls, not one combined query.

Handle pass. Trusted accounts, narrow window.

query: "latest on <topic> since:<yesterday>"
allowed_x_handles: ["growth_papa", "aleabitoreddit"]
from_date: <yesterday>

Keyword pass. Open field, reply noise removed.

query: "recent <topic> <ticker or key term> -filter:replies lang:en"
from_date: <yesterday>

Thread-fetch pass. Only when one post from the passes above needs context.

query: "<x.com URL of the post> replies and quotes"

Stop after three passes unless the user asks for more. If the handle pass and

the keyword pass disagree, report both rather than averaging them.

Reading results

Each result arrives as a URL plus a one-line summary, followed by a

Queries used: trailer that lists the exact queries the server ran. Keep the

trailer in your notes; it's the only provenance you have for what was actually

searched.

When you report:

• Cite the x.com URL for every claim drawn from a post.
• Label the source as an X post. A tweet is a signal that someone said

something, not a verified fact. Say "posted on X by @handle" rather than

stating the content as true.

• Quote sparingly. Summaries in your own words plus the link do the job.

Tandem with web_search

Run the same topic through web_search in the same turn. Then reconcile:

• Web-found event with no X echo: say so. It may be too fresh, too niche, or

the query missed it.

• X-only claim with no web coverage: say so, and flag it as unconfirmed.
• Both agree: cite both, X for the reaction, web for the record.

Don't let an X-only claim into a conclusion without that note.

Cost note

xAI bills X search at $5 per 1,000 calls. One x_search call is one search

turn. The three-pass recipe costs roughly $0.015. That's cheap, but it isn't

free, so don't loop on retries when the first pass came back empty; widen the

window once, then stop.

Backtest pointer

Query recipes and parity against other lanes are checked by

packages/omo-senpi/scripts/qa/x-search-backtest.mjs. If you change a rule in

this skill, update the query set there so the backtest still reflects the

recipe it's measuring.