Help

What this app is

Two products, one engine, one set of saved setups.

The tabs at the top left — Newsletter and Research — are two different questions, and everything below the tabs changes with them.

Newsletter

What happened this week, and what do we say about it? Scans the web on a topic you describe, drafts an editorial brief, and you choose what ships. Produces a brief.

Research

Who should we call, and why them? Finds companies, scores each one against criteria you write, and ranks them. Produces a report, exportable as Word, PowerPoint, PDF.

Both run the same three-pass engine and both save their setup as a preset you can re-run later. The difference that matters: a newsletter is prose you edit, a report is candidates you score.

Getting an account

There is no sign-up button, and that is deliberate.

Accounts are created for you by an administrator. Nobody can register themselves — this app reaches a live research budget and an outbound mailbox, so access is handed out rather than claimed.

1
Ask an administrator for an account

Ask whoever set this app up for your team — usually the person who sent you the link. Administrators are the only accounts that can add another one.

Tell them the name you want shown in the app and whether you need administrator rights (you probably do not — see below).
2
They create it under ⚙ → Manage users

They enter a username, your display name and an initial password, and hand you both. There is no invitation email: the credentials come to you however they normally reach you.
3
Sign in at /login

Username and password. The session is a cookie, so you stay signed in until you log out or the server restarts.

Changing or resetting your password is an administrator action. There is no "forgot password" link and no self-service change — if you are locked out or want a different password, ask an administrator. They set a new one on your account and tell you what it is.

Ordinary user vs administrator

Everyone can run both modes, review the output and download exports. An administrator additionally gets the ⚙ menu: managing presets, managing users, the intent options behind the newsletter setup screen, and appearance. Most people should not be administrators.

Access is removed by deactivating an account, never by deleting it — a deactivated user stops being able to sign in, and everything they created keeps their name on it. If your login suddenly stops working, that is the first thing to ask about.

Newsletter mode

Describe a topic → the app writes the prompts → generate → choose what ships.

1
Start at + New Article

The green button, top right, in Newsletter mode. Describe in plain language what the newsletter is about and who reads it. You are not writing a prompt — you are describing the job.
2
Answer two or three questions

Clarify reads what you wrote and asks the handful of things it genuinely cannot infer. Then Generate prompts turns your answers into the three prompts the engine will run: Scan, Writer, Editor.

These two buttons are the only things on the screen that call the API, and they are cheap. Nothing has been researched yet.
3
Review the prompts, then Generate

You land on the New Article screen with all three prompts filled in and collapsed. Open any of them to read or edit it. Beside each is a box where you can say what to change and press Regenerate prompt.

An edited-but-unsaved prompt is used once. To keep it, use Save as new preset.
4
Watch the run

Generate runs the three passes in the background and the run screen shows where it is — scanning, writing, recommending. Cancel genuinely stops the pass in flight; it does not just hide the screen. The scan is the slow, expensive pass.
5
Choose what ships — the review

The run lands in review, not in a finished newsletter. You get the scan table (everything found, with its relevance judgement) and the drafted pieces as cards. For each card you tick whether it ships, edit the text, and optionally attach a chart.

A chart needs an image URL that really resolves to an image, a caption, and an attribution link. You may attach one from any page — the rule that a source must have been scanned binds the model, not you.
6
Create

Create is the seam where a review becomes the document. Only the cards you ticked are assembled, cross-references to pieces you dropped are cleaned up, and the brief gets its title and date. Re-opening a created brief and pressing Create again is safe.

Afterwards: the brief page, an email preview at /briefs/{id}/email, and a PDF download.

Sending is manual, and gated. Creating a brief sends nothing. The ✉ Send button on a created brief (or a finished report) opens a panel: pick a list or tick contacts, see every name and address on a confirm step, then send — and each send is logged with who, when and to whom. Hard limits: a document with a citation flag refuses to send; a list marked external refuses entirely until chart captions are provably bound to their charts (people outside the firm get nothing until then); and no email of any kind goes out until the SMTP test on the Appearance page has passed. Contacts and lists live under ⚙ → Contacts & lists.

The Idea Bank

Every scan stores what it found, whether or not it made the newsletter. Idea Bank is that pile across all runs, grouped so a story seen repeatedly reads as one recurring idea. Hiding an item there removes it from the list without deleting it.

Research mode

Write criteria → run → read the ranking → export.

Research has no prompt-writing funnel, and that is the point: its configuration is the criteria, which you write directly and no model paraphrases. One form, one button.

1
Start at + New report
Describe what you are looking for and who it is for, then write the criteria. A new setup comes seeded with these 4 to edit, rather than a blank grid:
- hard Check size — How much the company is raising, in USD.
- hard Never — None of our hard exclusions apply: no crypto-native or token-based businesses, no pre-revenue companies, no consumer social, and no businesses whose primary revenue is advertising.
- soft Stage — The company is at a stage where our capital is the right shape.
- soft Timing — There is a reason to call them NOW rather than in six months.
Up to 12 in total — beyond that the scoring table stops being readable. Clear a row's statement to remove it.
2
Decide which criteria are hard

Hard means failing it disqualifies the candidate outright. Soft means it contributes to the score. Make something hard only if you would genuinely not take the meeting.
3
Use a numeric range where the answer is a number

A criterion can be a range instead of a statement — a minimum, a maximum and a unit. Then the app does the comparison rather than the model: the model reports the figure it found and the arithmetic decides pass or fail. Either bound can be left empty for open-ended.
4
Tick the formats you want — before you run

Word, PowerPoint, PDF. Each has an optional box for instructions specific to that format ("one slide each, top five only"). You can add the same note to more than one.

Formats are chosen now, not at download time. The write-up pass writes each requested format in the same call — so a format you did not tick was never written, and asking for it later is refused rather than exported empty. Getting it means a new run.
5
Save and Run are one button

Tick Save as preset and give it a name to keep the criteria for next time; leave it unticked for a one-off. Loading an existing preset copies its criteria into the form — edit freely, the saved one is untouched unless you tick and name it again.
6
Read the ranking
The review screen sorts everything found into groups, and nothing found is ever dropped:
- Ranked — cleared every hard criterion, scored, ordered. At most 10; fewer is normal and the screen says so rather than pretending otherwise.
- Disqualified — failed a hard criterion, shown with the reason.
- Considered, not ranked — cleared the hard criteria and did not make the cut, in the same order. This is the "next best" list.
Each candidate carries a score like 4/4 · 3 unknown. The fraction is soft criteria passed; the tail is how many could not be verified at all.

UNKNOWN is not a fail, and it is not a pass. A criterion the research could not evidence with a citation comes back unknown: it disqualifies nobody, it is left out of the fraction, and it costs the candidate rank. Read it as how much we can actually say about this company — a 3/3 with four unknowns is a thinner file than a 3/4 with none.
7
Export

Download the formats this report was written for, as many times as you like. Exporting costs nothing and calls nothing — it renders the stored report, so the same report always produces the same document.

Presets, archives and deleting

The parts that work the same in both modes.

Presets

A preset is a saved setup — the topic, the audience, and either the three prompts (newsletter) or the criteria (research). Presets are typed to their mode, so the Presets screen shows the ones belonging to the tab you came from. A newsletter preset and a research preset may share a name; two of the same kind may not.

Presets are retired, not deleted: a retired preset disappears from the "start from a preset" list and every brief or report that used it keeps pointing at it. Managing presets is an administrator screen.

Deleting

Nothing in this app is hard-deleted. Deleting a brief or a report hides it from its archive and it can be restored from the deleted list.

What a run costs

Real money, and the app keeps the receipt.

Every call to the model is recorded with its token counts and its cost, and each brief and report shows what it spent, pass by pass. The pattern in both modes is the same: the first pass — the web scan — is about 90% of the bill. The writing passes are rounding errors beside it.

Things that cost nothing, so you never need to hesitate over them: opening anything, re-downloading an export in any format, re-creating a brief, and loading or editing a preset. What costs money is Generate, Run, and the two setup buttons on the newsletter screen (a fraction of a cent).

If the footer says stub mode, the app is not calling the model at all — runs come back instantly with placeholder content and cost nothing. It is how the app is tested. Content produced in stub mode is not research.

Words this app uses

Most of these mean something narrower here than in ordinary use.

Preset: A saved setup you can re-run. Typed to its mode: a newsletter preset holds prompts, a research preset holds criteria.
Brief: What a newsletter run produces.
Report: What a research run produces. Holds candidates, their scores, and the write-up.
Candidate: One company found by a research run. Ranked, disqualified, or considered-not-ranked.
Criterion: One test a candidate is scored against. Hard (failing disqualifies) or soft (contributes to the score); a statement or a numeric range.
Pass: One call to the model. Both modes run three: find, then judge, then write.
Review: The state between a finished run and a finished document, where you decide what it says. Runs stop here on purpose.
Retire: Hide a preset from the pick-lists without breaking anything that used it.

When something looks wrong

I cannot sign in.: Passwords are reset by an administrator — there is no self-service reset. If it worked yesterday and does not today, ask whether your account was deactivated.
A run has been going a long time.: The scan pass is slow by nature — it is reading the web. The run screen shows which pass it is on. Cancel really does stop it.
The run finished but there is no document.: It is waiting in review for you to choose what ships. That is the design, not a failure — look for it in the archive with a review badge.
A research report will not export as PowerPoint.: That format was not ticked when the report ran, so the deck was never written. Formats are chosen before the run; re-run with it ticked.
Most criteria came back unknown.: The scoring pass may only cite sources the discovery pass actually read, so it will say unknown rather than guess. Usually the criteria are asking for something the coverage does not state — make them answerable from a funding announcement.
Nothing was found at all.: A short list is information. Check whether a hard criterion is narrower than you meant — a numeric range disqualifies silently and exactly.
The footer says stub mode.: The app is running offline with placeholder content. Nothing produced in that state is real research.

Anything else — a number that looks wrong, a document that came out malformed, a screen that errored — is worth reporting with the brief or report number. Both are in the URL.