BTC Live Dashboard
A weekly, Monday-morning Bitcoin briefing at bitcoin.jakepercy.com: live price and on-chain metrics, a probability-scored weekly analysis, a hard-money thesis, a strategy tab, curated headlines and global adoption tables, on one page.
It is written for three readers at once:
- Friends and family learning Bitcoin from scratch — the binding constraint. A sentence they cannot parse costs trust in the whole page.
- Intermediate observers staying current without news-ticker churn.
- Recruiters evaluating the author for a Bitcoin / blockchain role.
The editorial voice is calm and primary-source-driven. Every forecast carries an explicit probability and every forecast is scored: the Analysis tab shows the Brier calibration, and its Track record lists every resolved call, hit or miss. The site teaches time preference. It does not manufacture urgency, and it never gives financial advice.
Architecture
Vite + React 18 + recharts. No TypeScript, no state library, no backend. The page is static; live figures come from free public APIs in the browser.
index.html head, meta, and the #root marker the crawler rendition is injected into
src/
main.jsx mount
BitcoinDashboard.jsx the shell: header, tab bar, hash router, shared state
tabs/ one component per tab: Dashboard, Thesis, Strategy, Headlines, Adoption, Analysis
components/shared.jsx cards, Term (glossary popover), StaleNotice, …
hooks/ useBitcoinData (prices, mempool), useBitcoinNews (RSS via rss2json), ui
lib/ constants (API URLs, TABS), routes (hash scheme), format, market-math, glossary-match
data/ the content — the only directory the refresh automations edit
styles.css tokens (--elev-*, --fs-*) and the elevation material system
scripts/ build and check scripts (Node ESM, no dependencies)
automations/ one playbook per refresh skill, plus the analyst memory
.github/workflows/ pages.yml (build + publish to GitHub Pages on every push to master), ci.yml (pull-request checks)
src/data/ — content modules and their owners
| Module | Exports | Owner |
|---|
market-news.js | MARKET_THEME, CURATED_MARKET_UPDATES, FALLBACK_NEWS, NEWS_FEEDS | /refresh-market-news (weekly) |
headlines.js | HISTORICAL_HIGHLIGHTS, SIX_MONTH_HIGHLIGHTS | /refresh-market-news, /refresh-headlines-archive |
deep-analysis.js | DEEP_ANALYSIS — v2: probabilities, falsification conditions, calibration | /refresh-deep-analysis (weekly) |
forecast-ledger.js | FORECAST_LEDGER — the Track record | /refresh-deep-analysis (weekly) |
adoption.js | COMPANY_HOLDINGS, COUNTRY_HOLDINGS, ADOPTION_INDEX_TOP20, ADOPTION_DATA | /refresh-adoption |
fallbacks.js | FALLBACK_* constants, YARDSTICKS | /refresh-fallbacks (monthly) |
glossary.js | GLOSSARY | writers append with a source; the Editor proposes |
thesis.js, strategy.js | the static Thesis and Strategy tabs | hand-written; out of scope for the automations |
The app is split so the refresh skills never read or edit anything outside src/data/. SCOPE.md is the source of truth for what each skill may touch.
Routing
Tabs and in-page sections are addressable as #/<tab>/<anchor>, defined in one place: src/lib/routes.js.
#/thesis opens a tab; #/analysis/track-record opens a tab and scrolls to a section.- Tab slugs:
dashboard, thesis, strategy, headlines, adoption, analysis.
- Anchor ids live in
ANCHORS in the same file. npm run check:schema verifies they are unique and that every anchor a data module links to is defined there.
- A hash that does not start with
#/ is left to the browser, so plain in-page anchors keep working.
How the analysis is written
The weekly analysis, the headline cards and the adoption updates are written by AI agents (Claude models running the playbooks in automations/), not typed by hand. Jake Percy designed the pipeline and has refined it since May 2026 until the output reads the way he wants his own writing to read: plain enough for a newcomer, tied to primary sources, and open about what it got wrong. Each weekly update is committed and published by the Monday automation; the build and test gates run twice (locally before the push, then in pages.yml before anything is served), and he reviews the result on the live page.
What that refinement consists of, all of it in this repository:
- A written standard that code enforces.
SCOPE.md sets three reader tiers, word ceilings and the probability wording bands. npm run check:plain measures every rendered sentence against it, and a structural failure blocks the build.
- One agent per job. Fallbacks, adoption, market news, the headlines archive, the deep analysis and an Editor each run their own playbook (about 1,250 lines across
automations/*/README.md). The analyst and the Editor run on models pinned in .claude/agents/.
- An Editor that never sees the analyst's notes. After the writers finish,
/refresh-plain-language rewrites the rendered text for a newcomer without reading the analyst memory, so its shorthand cannot leak onto the page. A numeric fingerprint taken before the edit restores any figure the Editor changes.
- A cold-read test. A separate model sees only the rendered text and has to say what happened, why it matters to someone holding Bitcoin, and what would change the author's mind. Cards it cannot answer go back to the Editor once, and what still fails becomes a note for next week's writer.
- Forecasts that are scored, not just stated. Every watch-list item carries a probability. When its window closes it is scored hit or miss (partial credit was dropped on 2026-10-07) and the Brier score in the public Track record updates.
- Feedback that binds the next run.
automations/refresh-deep-analysis/memory/editorial-lessons.md and automations/refresh-plain-language/reader-feedback.md are Jake's to edit, and every writer treats a line there as a hard rule.
Nine of the 26 commits since the repository's July 2026 baseline change the agents' prompts or standards. The analyst's working memory is published as-is in automations/refresh-deep-analysis/memory/, misses included.
The weekly pipeline
/refresh-all is the orchestrator. It runs each refresh skill on its own cadence, each behind a significance gate that may close on a polled run with nothing material to say, then builds once at the end.
| Agent | Cadence | What it covers |
|---|
fallbacks | Monthly (30d) | Last-resort price / ATH / block / mined constants, YARDSTICKS |
adoption | Weekly (6d), Monday anchor | Public-company treasuries, sovereign reserves, adoption indices |
market-news | Weekly (6d), Monday anchor | Top-5 themes (Market Pulse + 4 cards), four-week highlights, RSS fallback |
headlines-archive | Monthly (25d) | Six-month archive of important headlines |
deep-analysis | Weekly (6d), Monday anchor, chains after market-news | Analysis tab synthesis, probabilities, Brier scoring, the forecast ledger |
Between Mondays, /refresh-all <hint> force-runs for genuinely major news; the trigger list is in automations/refresh-all/README.md. Force-runs never bump the cadence manifest, so the Monday baseline survives.
The Editor. After the writers and before the build, /refresh-plain-language reads every rendered field with fresh eyes and rewrites for a newcomer: consequence first, every term defined at first use, probability band phrases instead of numbers, no sentence over 45 words. It never adds a fact, never reads the analyst memory, and never touches a scoring field. npm run check:plain measures the result, and its exit code decides whether the Editor runs again, the build proceeds, or the file is restored from the pre-edit snapshot.
Memory and calibration. /refresh-deep-analysis keeps a project-scoped analyst memory in automations/refresh-deep-analysis/memory/: theme and actor files, a causal model, the watch-list ledger and calibration bins. Every watch-list row carries a probability on a 0.05 grid; when its window closes it is scored hit or miss against the threshold in its own wording (never partial; a forecast its wording cannot decide is marked no-resolution and left out of every score) and the Brier score updates. /consolidate-deep-analysis distills the journals each quarter.
The forecast ledger
src/data/forecast-ledger.js is the public, plain-English face of that memory: every resolved forecast with the date it was made, the odds given, what happened, its Brier score and one lesson, newest first — plus the aggregate score, the per-category bias and a calibration curve. It renders as Track record on the Analysis tab (#/analysis/track-record). The file is regenerated from the memory ledger, never hand-edited, and npm run check:schema enforces that the counts, the sort order and every row's arithmetic agree.
Build, checks and deploy
npm install
npm run dev # Vite dev server on :5173
npm run build # split build: dist/index.html + dist/assets/, crawler rendition injected
npm run build:single # one self-contained dist/index.html (CSS, JS and SEO content inlined)
npm run check:schema # shape gate for every src/data module — exit 1 on any violation
npm run check:plain # plain-language checker — 0 clean · 1 style · 2 Editor must fix · 3 structural, do not build
npm run check:llms # dist/llms.txt is present, valid, linked from the page and current — exit 1 on any problem; skipped without a build
npm test # check:schema, the glossary-matcher regression tests, then check:llms
scripts/render-seo.mjs regenerates a text rendition of the curated content on every build, so crawlers that never run JavaScript still read this week's analysis. After any src/data/*.js edit, rebuild.
Deployment is GitHub Pages. Every push to master runs .github/workflows/pages.yml: npm ci, the build, npm test, then publish dist/ to bitcoin.jakepercy.com. Weekly refreshes are committed on master by the Monday automation (/publish-weekly, see automations/publish-weekly/README.md) and ship on push; other changes arrive as pull requests, checked by ci.yml. The old addresses, jakepercy.com/bitcoin/ and nati-crypto.com/bitcoin/, redirect here.
llms.txt
The build also writes dist/llms.txt: a short plain-text guide to the page for AI readers, most of which run no JavaScript. It says what the page shows, how to read its stated probabilities and Brier-scored track record, what not to conclude from them, where the live data comes from, and the #/<tab>/<anchor> addresses. The page points at it from the <head> (index.html) and from the first line of the static text (render-seo.mjs).
- Where the prose lives.
scripts/llms.txt.tpl, hand-written. scripts/render-llms.mjs fills in everything that moves: tab names and routes (src/lib), endpoints and polling intervals (constants.js), the probability wording bands (SCOPE.md), the page's own stale-after rule (shared.jsx), and the weekly analysis and track record (src/data). The "Current state" block carries the analysis's own asOf date, never the build date.
- Quote check. Every double-quoted string in the template must still be on the page (
index.html, src/ or the static text in render-seo.mjs). Reword a label, a probability band, the Brier sentence or a disclaimer and the template needs the same change. The build only warns (lines prefixed llms:): the weekly refresh runs it unattended, and llms.txt is never allowed to be the reason it fails. npm run check:llms (and npm test) turns the same findings into failures.
- Contract. ASCII only, under 10 KB, and the opening blockquote must equal this dashboard's row in
content/llms/sites.json in the jakepercy.com repository (summary plus caveat). node scripts/check-llms.mjs in that repo validates every dashboard's file together; edit the sheet there, not here.
- See it without building.
node scripts/render-llms.mjs prints the file from the current source.
Contributing
- Read
CLAUDE.md, then SCOPE.md. SCOPE.md lists what each refresh skill may change, what is out of scope for all of them (the Thesis and Strategy tabs, live-API code, URL constants, everything under src/ outside src/data/), and the plain-language standard every sentence is held to.
- Plain language is a hard rule, not a preference. Run
npm run check:plain -- --file src/data/<file> on anything you write; fix hard findings once, re-run once, and list what is left.
- Edit data with narrow, anchored edits to the constant; never rewrite a data file wholesale. Re-sort the holdings tables by
btc descending after any row change.
- Primary sources only:
.gov, SEC EDGAR, official investor relations, bitcointreasuries.net, Farside Investors, BIPs. Major Bitcoin press as supporting context, never as the sole source.
- Design: no colored alert stripes or tinted callouts. Hierarchy comes from the
--elev-* elevation tokens in styles.css, and nothing renders below 11px.
dist/ is build output. Do not edit it.
Licence
Two licenses, both in LICENSE: the code is MIT; the analysis, forecasts, playbooks and analyst memory (src/data/, automations/) are © Jake Percy, all rights reserved, quotable briefly with credit. Nothing here is financial advice.
