Daily macro & markets read for Bangladesh banking professionals.
Numbers, news, and a banker's read on what matters. One brief. Every morning at 08:00 BDT. ~15 minutes.
Read it → thebrief.clauding-lab.com Subscribe (free) → thebrief.clauding-lab.com/#subscribe Install it → open the site on your phone → Share → Add to Home Screen (ships as an installable app since v2.3.0)
The "1c" identity: light mode is one uninterrupted paper sheet; dark mode is the full-bleed ink band. A light/dark toggle whose choice persists per device, the compact editorial scale, and a six-cell market snapshot strip on every issue.
| Light | Dark |
|---|---|
 |
 |
The Brief is a daily editorial brief that synthesises Bangladesh's macro and markets data into a single ~15-minute read for senior banking professionals — business heads across corporate, SME, and retail, plus risk heads, plus treasury heads, at Tier-1 banks in Bangladesh.
Every morning at 08:00 BDT, a Python pipeline pulls fresh data from EconDelta (the upstream scraper system) via a Supabase store, hands it to Claude (Anthropic's LLM) running with a tightly-scoped Desk Editor prompt, and publishes the finished brief — structured prose + verdicts + numbers + headlines + chart data — back to Supabase. A Next.js single-page app on Vercel reads the latest issue and renders it as a morning-paper page — one steel-paper sheet in light, a full-bleed ink masthead band in dark — in the reader's choice of theme. Subscribers receive an HTML + plain-text email at the same moment via Brevo.
There is no human in the loop on the daily run. The brief writes itself, reviews itself (a subeditor LLM pass), and then fact-checks its own prose deterministically — every figure in the copy must trace to a builder value, a machine-fed history fact, or a documented derivation, and an invented market statistic in the DSE section hard-holds the publish rather than shipping. The voice is consistent because the prompts are fixed; the data is fresh because the upstream scrapers finish about five hours earlier.
Senior banking professionals at Tier-1 banks of Bangladesh who need:
- A defensible morning read on the macro arc — Bangladesh Bank policy, FX, sovereign debt, fiscal pressure
- Honest framing of the markets day — rates curve, DSEX, T-bills / T-bonds, gold, Brent
- A weekly wrap on Friday that synthesises the 5-day arc and flags next week's calendar items
- A single source that doesn't pretend daily volatility is a trend
The brief assumes the reader already knows what NPL means, what CAR is, and how to read a yield curve. It does not explain finance basics. It does explain what the numbers imply for desk positioning.
Each daily brief contains, in order:
| Block | What it is |
|---|---|
| Masthead | The identity band (paper in light, ink in dark): issue number, date, lens pill, publish stamp, theme toggle, "Today's Call" — the editorial thesis |
| Snapshot strip | 6 KPI cells (USD/BDT, DSEX, 91-d T-Bill, Brent, Gold, monthly remittance) with sparklines — derived live from section metrics |
| Banking | NPL, CAR, defaulted-loan stock, sector credit growth |
| Markets | Rates curve, DSEX, FX flows, T-bill / T-bond yields |
| Real Economy | Macro indicators, BB monetary, fiscal items |
| Policy | NBR, regulatory, budget, IMF items |
| Banker's read | 2-3 sentences per major section: stance, watch list, risk vectors |
On Friday, the cadence shifts — Friday issues are weekly wraps: a 5-day synthesis with a "biggest movers" paragraph and a watch list for the next week.
| Day | Fires at | Type | Lens |
|---|---|---|---|
| Mon | 08:00 BDT | Daily | data-driven (FX-runway / credit-cycle / rates-curve / sovereign-debt / external-shock) |
| Tue | 08:00 BDT | Daily | data-driven |
| Wed | 08:00 BDT | Daily | data-driven |
| Thu | 08:00 BDT | Daily | data-driven |
| Fri | 08:00 BDT | Weekly wrap | 5-day synthesis |
| Sat | 08:00 BDT | Daily | data-driven |
| Sun | 08:00 BDT | Daily | data-driven |
Brief.timer on Hetzner runs Mon..Sun 02:00 UTC — every day, Saturday included.
┌─────────────────────┐ ┌───────────────────────┐ ┌──────────────────────┐
│ EconDelta scrapers │ │ Supabase │ │ The Brief │
│ (ExonVPS) │ ────▶ │ metric_history, │ ────▶ │ pipeline (Hetzner) │
│ aggregate 02:55 BDT│ │ briefs, sections, │ │ fires 08:00 BDT │
│ daily │ │ news, subscribers │ │ │
└─────────────────────┘ └───────────────────────┘ └──────────┬───────────┘
│ │
│ │ writes brief
│ ▼
┌──────────┴───────────┐ ┌──────────────────────┐
│ Next.js SPA │ │ Brevo │
│ (Vercel) │ ◀────── │ email API │
│ reads briefs │ │ sends to all rows │
│ via RPC │ │ in subscribers │
└──────────────────────┘ └──────────────────────┘
- EconDelta scrapers — separate repo (
clauding-lab/econdelta), runs on ExonVPS. Fetches from ~01:30 BDT and finishes writing to Supabasemetric_historywhen its aggregate stage lands at ~02:55 BDT — deliberately ahead of this pipeline's 08:00 fire, so a brief reads the SAME morning's data rather than yesterday afternoon's. - The Brief pipeline — Python, runs on Hetzner via
brief.servicesystemd unit. Fires at 08:00 BDT. Six stages:- Gather — per-section builders pull from
metric_history(Supabase) andlatest.json(rsync'd snapshot from ExonVPS) - Pick lens — data-driven (NPL stress → credit-cycle; Brent spike → external-shock; etc.) or weekly_wrap on Friday
- Editor LLM — Claude with
editor_v6.txt/editor_v6_friday.txtproduces structured JSON (Today's Call, section verdicts, banker_read blocks, hero selection, etc.) - Subeditor LLM + deterministic gates — a second Claude call reviews the editor's output against the Pydantic schema + editorial rules (a malformed or incomplete review is retried once, then HOLDS the publish rather than shipping unreviewed), followed by a hard-coded fact-checker: every currency/percent/bp figure in the prose must trace to a builder value within tolerance, machine-computed disclosures (import-cover periods, the Real Policy Rate's own arithmetic) are stamped onto the page deterministically, and a fabricated count claim — a "ten-session low" no data supplied — warns everywhere and hard-fails the publish in the DSE section, where the pipeline computes the true rank itself
- Publish —
v6_publisher.publish_brief()clears any same-issue row, then writes the brief as adraft, POSTs sections/metrics/news/chart_series, and flipsstatus='published'as the LAST call — a two-phase write, not a single DELETE + INSERT (idempotent for same-day re-runs). A reconciliation pass force-reinserts a short list of protected metrics (currently the BB policy corridor) if the editor dropped them, and hard-fails the publish if one is still missing - Notify —
notifier.notify()fetches all rows fromsubscribers, renders an HTML + plain-text digest, and POSTs ONE Brevo call per subscriber (never a multi-recipientto:list — that would leak every recipient's address to the others). Fail-open — never crashes the publish.
- Gather — per-section builders pull from
- Supabase — the data layer. Tables:
metric_history(indicator time series),briefs,sections,metrics,news,chart_series,chart_notes,subscribers. Service-role auth for the publisher and the subscribe form; anon read-only auth for the SPA. - Next.js SPA —
app/, deployed on Vercel. Reads the latest brief via theget_latest_briefSupabase RPC and renders it in the "1c" editorial identity: one steel-paper sheet in light, a full-bleed ink masthead band in dark, themes the reader flips (charts rebuild on flip), compact type scale, KPI tile grids, an installable PWA manifest, and a print contract that produces a clean light PDF from either theme. Hash-routed scroll-spy navigation, IntersectionObserver-driven sticky bar, no-JS-needed first paint via SSR.
| Layer | Tool |
|---|---|
| Editorial pipeline | Python 3.11+ (stdlib + anthropic + pydantic) |
| LLM | Anthropic Claude (single model pin: Opus 4.8 via Claude CLI — no Sonnet path) |
| Data layer | Supabase Postgres + PostgREST |
| Web app | Next.js 16 (App Router) + React 19 + TypeScript |
| Charts | Chart.js 4 + chartjs-adapter-date-fns |
| Hosting (compute) | Hetzner (clauding-lab, Ubuntu, systemd) |
| Hosting (web) | Vercel |
| Brevo (transactional API, one POST per subscriber) | |
| Scrapers (upstream) | clauding-lab/econdelta on ExonVPS |
Subscribe: thebrief.clauding-lab.com/#subscribe — name, organisation, email, one click. No tracking. No marketing list.
Unsubscribe: reply to any brief email with "Unsubscribe" in the subject. (One-click automation is on the roadmap; manual reply works today.)
the-brief/
├── app/ # Next.js App Router — the SPA
│ ├── layout.tsx # Root layout, metadata, PWA config
│ ├── page.tsx # Server-side initial fetch via Supabase RPC
│ ├── components/ # React components (Masthead, Section, SnapshotStrip, SecNav, ThemeToggle, …)
│ └── globals.css # Design tokens (light/dark steel-crimson; band tokens per-theme) + print contract
├── brief/ # Python pipeline
│ ├── cli.py # Entry: `python -m brief.cli run --publish`
│ ├── pipeline.py # Section-gather orchestration
│ ├── pipeline_v6.py # V6 publish flow (editor + subeditor + retries)
│ ├── v6_publisher.py # PostgREST writer
│ ├── v6_schema.py # Pydantic schema for the editor's JSON output
│ ├── notifier.py # Release email notifier (Brevo)
│ ├── builders/ # Per-section data gatherers
│ ├── claude/prompts/ # Editor + subeditor LLM prompts
│ ├── headlines.py # News scraper
│ └── history.py # PostgREST history client
├── lib/ # Shared SPA helpers (format, fallback, supabase client)
├── deploy/ # systemd unit, env example, install/uninstall scripts
│ ├── brief.service # The daily (Mon–Sun) timer-driven oneshot unit
│ ├── brief.timer # OnCalendar=Mon..Sun 02:00 UTC
│ └── brief.env.example # Required env vars
├── docs/ # Design docs, plans, ops notes
│ └── superpowers/specs/ # Per-feature design specs
├── tests/ # pytest suite (1,100+ tests as of v2.3.1)
├── migrations/ # SQL migrations against Supabase
├── public/ # Static assets
├── CHANGELOG.md # Release history
└── README.md # This file
# 1. Clone + create venv
git clone https://github.com/clauding-lab/the-brief.git
cd the-brief
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt -r requirements-dev.txt
# 2. Provide env vars (the project's /etc/brief.env on Hetzner is the canonical source)
cp deploy/brief.env.example /tmp/brief.env
# Fill in SUPABASE_URL, SUPABASE_SERVICE_ROLE_KEY, BREVO_API_KEY, FROM_EMAIL, …
# 3. Run pytest
.venv/bin/pytest -q
# 4. Run the pipeline locally (dry-run = no Supabase write)
set -a && source /tmp/brief.env && set +a
.venv/bin/python -m brief.cli run --publish --dry-run --no-notifyCLI flags:
--publish— V6 publish flow (writes to Supabase)--dry-run— go through the pipeline but skip the Supabase write--no-notify— skip the subscriber email after a successful publish--today=YYYY-MM-DD— override the date (useful for re-runs)--write-fixture=<path>— (with--dry-run) write the finished brief as JSON instead of publishing; no Supabase write, no email--preview-notify— (with--write-fixture) best-effort ping to Discord + email with the preview link
Exit codes: 0 ok · 1 error · 3 dry-run ok · 4 publish failed
brief.cli run --publish --dry-run --write-fixture=public/fixtures/<name>.json runs the full pipeline (editor + sub-editor) and writes the result to a fixture file, without touching Supabase or sending any email. Open /preview?fixture=<name>.json on a Vercel branch preview (or locally) to see the brief rendered exactly as a reader would, behind a yellow "PREVIEW MODE" banner. This is the safe way to check an editor-prompt or pipeline change before it reaches production.
# Inside the same checkout
npm install
# Provide env vars (or .env.local)
echo "NEXT_PUBLIC_SUPABASE_URL=https://<project>.supabase.co" > .env.local
echo "NEXT_PUBLIC_SUPABASE_ANON_KEY=eyJ..." >> .env.local
# Dev server
npm run dev # http://localhost:3000
# Production build + typecheck
npm run build
npx tsc --noEmit
# Lint
npm run lint| Where | Runs | Cadence | Files |
|---|---|---|---|
Hetzner (clauding-lab) |
brief.service — Python pipeline |
Daily · 02:00 UTC | /etc/systemd/system/brief.{service,timer}, /etc/brief.env |
| Vercel | Next.js SPA | continuous deploy from main |
app/, lib/, public/ |
| Supabase | Postgres + RPC | continuous | migrations/, RPC get_latest_brief |
| ExonVPS | EconDelta scrapers (upstream) | daily ~05:30 BDT | separate repo |
| Brevo | Transactional email send | per-publish | API call from brief.notifier |
# Systemd service log on Hetzner (tail)
ssh adnan@<hetzner> "tail -50 ~/the-brief/logs/brief-systemd.log"
# journalctl view
ssh adnan@<hetzner> "journalctl -u brief.service --since '12h ago' --no-pager"
# Vercel deploy + runtime logs
vercel logs --follow # Vercel CLI# Re-run a publish on Hetzner manually (e.g. after a missed auto-fire)
ssh adnan@<hetzner> 'cd ~/the-brief && \
LOG=logs/manual-run-$(date -u +%Y%m%d-%H%M%S).log && \
nohup bash -c "set -a; source /etc/brief.env; set +a; \
.venv/bin/python -m brief.cli run --publish" > $LOG 2>&1 < /dev/null & \
disown; echo "LOG=$LOG"; echo "PID=$!"'The Brief uses semantic versioning. Releases are tagged on main with v<MAJOR>.<MINOR>.<PATCH> and published as GitHub releases with notes.
- Major — schema break in
briefs/sections, or pipeline architecture change - Minor — new section, new editorial feature, new visible UI surface
- Patch — fixes, copy changes, deferred-cleanup commits
See CHANGELOG.md for the full release history. Current: v2.5.0 — the honest-clock release (three-month yield ladder with its auction footnote, float tails cut, reserves on a monthly clock, clean dependency audit on Next.js 16.3.4) on top of v2.4.0's paper-light release.
- EconDelta — upstream scrapers that populate
metric_historyin Supabase. Without EconDelta the brief has no data.
Copyright © 2026 The Brief — all rights reserved. The Brief is a private editorial project; the code is published for transparency, not for redistribution.
Issues, questions, or polish suggestions: open a GitHub issue or reply to any brief email.