How a brief is produced.

The data layer

Every brief starts with a paid call to the x402 research agent, a full-archive X data service. Payment is on-chain via BSV micropayment — no API subscription, no enterprise data contract, no account to provision. The call returns the top 100 posts by engagement matching the query, over a rolling 30-day window, filtered to original posts (no retweets) above an engagement floor.

The engagement filter exists because on a topic with 10,000 posts in 30 days, most of that volume is noise: bots, quote-farming, low-effort takes. The top 100 by engagement is usually where the human signal lives.

The synthesis layer

For self-serve briefs, Claude Haiku turns a customer's topic into a structured X search query, then Claude Opus 4.7 reads the full returned corpus and writes the brief. For Custom and Deep dive briefs, the query and synthesis prompt are hand-designed for the specific decision before any model runs.

The deliverability check

Before Opus writes anything, it evaluates whether the corpus actually answers the customer's question. If the corpus is mostly off-topic (keyword collisions matching unrelated meanings), dominated by a single account, or too sparse for honest generalization, Opus refuses to write a brief. The customer gets a rerun offer at no extra charge instead.

We'd rather skip the sale than ship a misleading brief.

What the brief contains

If the corpus passes the deliverability check, Opus produces the brief in six sections, in this order:

1. Corpus shape.Post count, date range, dominant voices, who's notably absent, and what proportion of the corpus is on-topic versus tangential.
2. The dominant narrative. What most of the corpus is actually saying, and who is driving it (named accounts).
3. Supporting threads. Secondary arguments, specific entities, named products or programs, key dates and events surfaced in the corpus.
4. Noise to discount. Promotional spam, listicle reposts, hype accounts, off-topic matches. Named, not belabored.
5. Voices and sources to follow. The three to five accounts worth tracking for someone monitoring this topic going forward.
6. Three observations.What appears stable, what's still being figured out, what to watch over the next two to four weeks.

Limited-signal handling

When the corpus is real but thin, the brief opens with an explicit "Limited-signal note" and uses sample-aware language throughout — "in this sample," "the visible discussion suggests," "among the posts reviewed" — instead of market-consensus claims like "X users believe." Findings are framed as directional signal from a specific sample, not as broad consensus.

Sourcing rules

• Named accounts are preferred over aggregated claims.
• Every hard claim ties back to a specific post in the corpus.
• The full corpus — every post sorted by engagement, with handles and source URLs — is preserved at a permalink in every brief. You can audit any claim against the raw data.
• Style is short and direct. No AI-flavored softening. No em-dashes between clauses (a tell of generated prose).

What you receive

A PDF brief at the email you provided, with a permalink to the raw corpus that fed it. Bookmark the link to return to the underlying data anytime. Briefs are preserved indefinitely.