Every claim is a checkable process —see how a morning runs
What we read · and why
Four source classes. Zero social feeds.
About fifteen machine-readable feeds — we read only what publishers offer machines, never behind a login or paywall, and never scraped from social or video platforms. Most popular crypto content is a retelling of 5–10 originals; we read the originals.
- Source 01
Diverse crypto press
Mainstream outlets to professional sceptics — one event from many angles, so no single outlet’s slant carries a story.
How · public feeds - Source 02
Business & markets press
Crypto is priced in Washington, at the Fed and in the bond market — not only on-chain. So we read the macro desks too.
How · public feeds - Source 03
Primary sources
Central-bank and statistical releases: the origin itself, not the third-hand retelling everyone else recycles.
How · official releases - Source 04
On-chain & market data
The few numbers your theses actually ride on — read straight from the chain and the market, not from someone’s take on them.
How · data APIs - Signal only
A breadth signal
“What is everyone talking about?” — a completeness check and a ranking input. Read as a signal only, never as the text of a story.
How · ranked, never quoted
The pipeline · end to end
Seven steps, every run.
Half the work takes no AI at all — collecting, deciding and checking are deterministic. AI only writes and cross-checks. And no briefing ships until it passes both gates and a human has read and released it.
Seven steps · every run
Collect
Deterministic · the day’s items, run or wait
Write
AI · clusters, ranks, writes both sides
Validate
Deterministic · structure, sources, numbers
Cross-check
Independent AI · every claim against its sources
Release
A human · reads it and releases it
Publish
Deterministic · only what was released
RepairBack to the writer
Step 01
deterministic
Collect
Pull the day’s items from the source feeds and on-chain data. No logins, no paywalls, opt-outs respected.
Step 02
deterministic
Decide
Is there enough new signal to publish? A sharp market move can force a run — but never a briefing for its own sake.
Step 03
AI
Write
Cluster what independent outlets carry, rank by long-term relevance and your theses, and write the summary plus the strongest bull and bear case — in our own words.
Step 04
deterministic
Validate
A strict gate: structure, sources present, every number found in the reporting, and no copied wording.
Step 05
independent AI
Cross-check
A second, independent model verifies every claim against its sources. What either check flags goes back to the writer for repair; what still cannot be verified is dropped. A run left with too few verified stories is not published.
Step 06
a human
Release
A human reads the finished run — every story, both sides, the source line — and releases it with a tick. Or discards it. No briefing appears without the tick.
Step 07
deterministic
Publish
Atomically, at the moment of release. Until a new run is checked and released, the app keeps showing the last briefing. Rather stale than wrong.
The Moonkelp engine · every morning
The morning, on a loop.
What the engine does before you wake up — collected, ranked, written, checked. Then it waits: no briefing is published until a human has read and released it. Stylised and endless; the real thing, replayed line by line exactly as it ran, is one click away.
A stylised loop of the engine’s seven steps: collect, decide, write, validate, cross-check, release, publish.
The method · point by point
Eight points, in full.
Moonkelp is written by machines and released by a human before you read it. That only works if the machine’s rules are public. This page is the whole method — a three-minute read, kept current.
01
the product
What Moonkelp is
Moonkelp is a daily 60-second briefing for long-term crypto investors: the reports that matter across crypto and macro, the strongest bull and bear case for each, filtered by the theses you hold — threats first. It is built to answer one question —am I fully informed? — and then let you go. Target reading time: about a minute.
02
sources
Where stories come from
We read the source layer, not the echo. Around fifteen sources across four classes: editorially diverse crypto press (from mainstream to professional sceptics), business and markets press,primary sources(central-bank and statistical releases), and raw on-chain data for the indicators. We read only what publishers offer for machine reading, and we respect opt-outs. Most popular crypto content is a retelling of a handful of originals — reading the originals is the entire point.
03
ranking
How ranking works
A story rises when severalindependentoutlets carry it, when it matters on a multi-year horizon, and when it touches the investment theses our readers follow — threats before comfort. A story never rises because it would generate clicks. “Bitcoin moves 3 %” is not a top story.
04
both sides
Both sides, by design
Every story carries the strongest honest bull case and the strongest honest bear case — steelman arguments, written from the sources, with links back to the originals. If a side has no grounded argument, we say so rather than invent one.Perspectives, not investment advice.
05
the gates
Checked, released, then published
Every briefing passes two machine gates before it is even finished: a deterministic validator (structure, sources, numbers, no copied wording) and a second, independent AI model that cross-checks every story against its sources. And then comesa human— who reads the finished run and releases it with a tick. Without that release nothing appears, and there is no deadline after which a briefing goes live on its own. A run that fails a check, or that is not released, never appears; the app keeps showing the last briefing.Rather stale than wrong.
Anthropic, Claude, Google and Gemini are trademarks of their respective owners. Moonkelp is neither affiliated with nor endorsed by them. The models in use can change; what does not change is that the two come from different providers.
06
never
What we never do
- No scraping of social platforms or video sites, no paywalled content.
- No copied sentences — our own words only, with links to the originals.
- No ads, no sponsored placements, no engagement optimisation.
- No buy or sell recommendations, no price targets.
- No feed, no engagement loop, no design that wants you to stay. The one notification is opt-in and arrives once a day.
07
ai, labeled
AI, labeled
The texts are machine-written and marked as AI-generated — written by a machine, released by a human. We keep the label even though content under human editorial control would be exempt from it: the texts come out of a model, and that belongs on the tin. The sources are linked on every story, every story names the thesis it speaks to, and each briefing carries the number of claims that were checked and the number that were dropped — the failures are printed, not hidden.
No check prevents every mistake. What happens when one gets through anyway is set out in the corrections policy — including how to report it, within what deadline, and how the correction is marked on the story.
08
fine print
The fine print
Research and perspectives, not investment advice. Texts are AI-generated from public reporting, cross-checked by a second model, released by a human before publication, and linked to their sources. Nothing here is a recommendation to buy or sell anything. Questions about the method: see theimprintfor contact.