Stablecoin Pulse Report

A concise, actionable monthly read of the stablecoin market: page 1 is the pulse (KPIs, TL;DR, top movers with their causes, launches, depegs and incidents, market-wide signals, what to watch); page 2 is the evidence annex with linked sources. The value is the why: "PYUSD +25%" is noise; "PYUSD +25%, half of it from its Aave/Arc integration" is the product.

Read editorial-rules.md before writing copy, research-briefs.md before dispatching research, and layout.md before building.

Inputs

• Window end: the fetch time (UTC), with a rolling 30-day window matching Pharos circulatingPrevMonth. The fetcher has no historical-end option; run it on the last days of the month or the 1st of the next, and use data/market.json.window for the actual dates.
• Month label (e.g. October 2026) and partner name for the footer (default Polaris).
• Work directory: agents/<YYYY-MM-DD>-pulse-report/ (ignored scratch). Prior months live in sibling agents/*-pulse-report/ directories; reuse the previous content.json as the structural starting point, never its facts.

Prerequisites: PHAROS_API_KEY in the process environment or root .env.local (environment wins; report the name only if missing); Playwright Chromium or Google Chrome; Poppler (pdfinfo, pdftoppm) optional. X research needs live X search (see research-briefs.md).

Workflow

1. Pull data

Use the supported repository Node runtime (>=24.16 <27) with workspace dependencies installed, including tsx; the fetcher imports canonical shared TypeScript supply helpers.

node --import tsx .codex/skills/stablecoin-pulse-report/scripts/fetch-pulse-data.mjs --out agents/<YYYY-MM-DD>-pulse-report

Read data/brief.md end to end, then the data traps in editorial-rules.md. Before any research, list the candidate stories: the top gainers and losers by USD, material percentage movers, chain shifts, flagged single mints or burns (large-mints.json), unit-versus-price splits (unit-moves.json), and depegs. Resolve identity collisions (same ticker, different issuer) now.

2. Research in one parallel batch

Split the candidate stories into the five clusters in research-briefs.md. When delegation is available, dispatch one web researcher and one X scout per cluster together in a single batch (ten agents), all reading data/brief.md. Preflight X search with a control query first; if unavailable, run web-only and state "X unavailable" in the handoff. Without delegation, perform the research and skeptical review sequentially and disclose that review was not independent. Save every result verbatim to research/<Agent>.md (or named sequential-pass notes).

3. Write content.json

The orchestrator writes the copy; do not delegate synthesis. Follow the schema and character budgets in layout.md and the evidence ladder in editorial-rules.md. Every number must trace to data/ or a cited source; every causal claim must match its evidence strength.

4. Fact-check

Dispatch one independent read-only reviewer when available (prompt in research-briefs.md) over content.json, data/, and research/; otherwise use the disclosed sequential review. Apply every accepted fix yourself; re-verify disputed numbers against data/ rather than choosing between agents.

5. Build and inspect

Run the layout build command from the repository root.

The build fails on overflow or a page count other than two. Open out/page-1.png and out/page-2.png and inspect them visually: no clipping, collisions, split −$ amounts, or orphaned heading words. Fix fit by trimming copy, never by shrinking type below the documented minimum.

Completion Report

Report the PDF path, the three to five headline findings, drivers that remain unexplained ("no source names…"), data caveats that shaped the copy, the fact-check verdict and fixes applied, X search spend if known, and which agents or models ran. Do not commit or publish the report unless asked; it lives in ignored scratch.