Settings

Your Profile sits in the sidebar on its own, because it is working data rather than configuration. Settings is what's left: your sign-in, the keys research uses, how batches run, and the rules your emails follow.

8.1 Your profile#

Full screen

The same information as guided setup, in one page, plus a completeness indicator that names the specific fields still missing rather than showing you a bare percentage.

This is the most consequential page in the app. Every match is scored against it and every draft is written from it, so a thin profile produces thin results everywhere downstream.

Come back and update it when:

• Your research interests shift. Even a small change to your topic list moves scores.
• You finish a project or a paper worth citing.
• You decide funding is now non-negotiable, or you rule out a region.
• A draft says something about you that isn't quite right — fix the profile, not just the draft.

Guided setup at the top right walks you through the same fields in steps, pre-filled with what you already have. It's the friendlier way to do a big update.

Matches record which version of your profile they were scored against. After a significant rewrite, put your important professors back through a batch so their scores reflect the current you.

Where this is going#

Three fields near the bottom of the profile describe direction rather than history, and they carry more weight in drafts than their size suggests. A professor reading a cold email is partly asking "is this person heading somewhere my lab helps with", and nothing else on the profile answers that.

• After the degree — academia, industry, either, or not saying yet.
• Next few years — what you want to be working on by the end of the programme.
• Longer term — where it leads: a lab of your own, a particular problem, a field you want to move.

"Not saying yet" is a real answer, offered deliberately rather than as a placeholder. A year out from applying, undecided is often the honest position, and a draft that hedges truthfully reads better than one that asserts a five-year plan you do not have. What you should avoid is leaving it blank and leaving the two prose fields empty — then a draft has nothing to work with and either says nothing about your direction or reaches for a generic ambition.

Keep both prose fields short and concrete. "Make robot manipulation work in unstructured homes" is usable; "contribute to cutting-edge research" is not, and will be dropped by the writing rules as filler.

8.2 Settings → Account & Security#

• Email address. Change the address you sign in with. You'll need to confirm the new one from your inbox before it takes effect.
• Password. Set a new one. Choose something you don't reuse.
• Two-factor authentication. Scan the QR code with any authenticator app — or type the key in manually — then enter the six-digit code to switch it on. Worth doing: this workspace eventually holds your offers.
• Sessions. Sign out everywhere. Use it if you've signed in on a machine you don't control, such as a lab computer.

8.3 Settings → Research Access#

Research runs outside the app — in an AI assistant on your own machine, or in Claude on the web — and it needs either a sign-in or a key to write into your workspace. Chapter 10 is the full walkthrough, with a page for each assistant; this section is the reference.

Full screen

Connect an assistant, at the top of the page, writes the setup for you with your endpoint and key already filled in. Copy the block for whichever assistant you use rather than assembling it by hand.

Issue two keys, not one. One with Submit research, one with Check research. Whatever submits a finding is refused when it tries to confirm that same finding, and that independence is the only real check in the pipeline — with a single key, research runs and then stops dead at verification.

1. Name the key after whatever will use it — "laptop research assistant" beats "key 1" when you're looking at a list six months later.
2. Tick only the permissions it needs:

1. Click Issue key and copy the key immediately. It's shown once and can't be shown again.

The table below lists your keys, the first few characters of each, their permissions, when each was last used, and whether it's still active. Revoke kills a key instantly.

Assistants that signed in appear in this same table — there is no separate "devices" or "connected apps" list anywhere in settings. When you connect Claude, Codex or anything else by signing in rather than by pasting a key, the connection is added here as a row carrying a Signed in badge, labelled with the name the assistant registered under (so "Claude", not a name you chose). It has permissions, a Last used time and a Revoke button like any other row, and it counts as a connection for the two-connection rule below. Signed-in connections also expire on their own, and show as Expired when they do — reconnecting in the assistant creates a fresh row.

So an empty table genuinely means nothing is connected. If you signed an assistant in and see no row, the sign-in did not complete, or it completed against a different account — not a display problem.

Sessions in Account is a different thing: it signs you out of the web app. It does not touch these connections, and revoking here does not sign you out.

Housekeeping worth doing: revoke keys you no longer use, and revoke anything whose Last used is long past. No permission here can send email.

8.4 Settings → Research Batches#

How much research runs at a time, how often, and how much of it you review. Covered in detail in chapter 3.

Note that How much do you review? decides where checked findings land. Anything other than "publish once checked" sends them to Review → Ready to publish first, and nothing downstream can cite them until you publish — see §3.4a.

This page also holds the two lists of tick boxes that decide which kinds of people are worth researching — appointment type and seniority. They are defaults rather than filters: any professor can be switched on or off individually from the Professors list, and nobody is skipped before an agent has established what they actually are. Full explanation in §3.2a.

8.4a Settings → Model Routing#

Research is not one job. Tidying a list of links is mechanical; judging whether your work genuinely overlaps a professor's is not. Sending both to the same model means either paying top rates to normalize URLs or handing the match analysis to something too cheap to do it — and the second failure is the dangerous one, because it doesn't produce an error. It produces a fluent paragraph with the honest mismatches quietly missing.

So each stage has a model tier, and this page shows them:

The tiers describe the work, not a particular company's models — the same three apply whether you're running Claude, Codex or Gemini.

Override any of them. A floor and a ceiling let you say "never use the cheapest" or "never use the most expensive" without editing every stage.

Which assistant, and which models#

Which assistant are you running? tells the platform whose models to name at each tier. It doesn't switch anything on its own — Claude Code and Codex each pick their own model; this tells them which one to ask for. Leave it on Whatever I'm connected with unless you have a reason.

Each tier lists the models to try, best first. The second one is used only when a model is unavailable — rate-limited, not enabled on your account, or retired — never because an answer looked wrong. That distinction matters: a chain that fell back on quality would quietly move your match analysis to a smaller model, which is the failure this whole page exists to prevent.

You can type any model id, including one this build has never heard of. The list is a default, not a permitted set — if your account has early access to something, name it. Reset puts a tier back to the default.

8.5 Settings → Writing Rules#

Full screen

Two documents shape every draft, and they do not cover the same ground:

• Email template — the structure your outreach follows: what the opening does, what the middle proves, how it closes. Email only.
• Humanizer — how it must be written. This is what keeps drafts sounding like a person rather than a form letter, and it's where to put your own preferences: banned phrases, length limits, how you sign off. This one is not email-only. Statements of purpose, research proposals, cover letters, the why paragraphs — everything an admissions committee reads is written under it and checked against it.

Rules for one thing at a time#

The base humanizer applies to everything, which is also its limit: what makes a CV line good is not what makes a why this professor paragraph good, and one document that tries to say both ends up saying neither. So the Per-field rules tab holds a third kind of document — your own rules for a single field, applied on top of the base.

There are twenty-two fields, grouped by what they are: the two outreach strategies, the seven collaboration sections, the document kinds, and the CV import. Each one lists what is already checked for it before you write anything at all — a funding note needs a figure in it, a CV line has to start with what you did, a why this university argument has to name something that is only true there. Most fields need nothing added, and a field marked Built-in checks only is not misconfigured.

Where your field rules and the base disagree, the field rules win. That is deliberate, and it is why the two are kept as separate documents rather than merged: a rule you wrote for one surface should not quietly become a rule everywhere.

Edit either one and click Publish new version. It takes effect immediately for drafts written afterwards; existing drafts are untouched and each records the version it followed, so you can tell what changed and when. Version history lists every version, and you can reactivate an older one if a change made things worse.

Small, specific edits beat rewrites. If drafts open too warmly for your field, add one rule about openings rather than rebuilding the template.

Tone samples — teaching it your voice#

Rules describe how to write. A tone sample shows it. Below the two documents on the same page is a panel where you paste writing and the app measures it, then holds drafts to what it measured rather than to a generic idea of "professional".

Give each sample a short name (Thesis introduction), paste a few paragraphs, and say where it came from — the source is the whole point, because the three are used differently:

Samples under 60 words are rejected. Below that the measurements are noise: one long sentence in a fifty-word passage moves the average further than any real habit would.

What gets measured is deliberately small and checkable — average sentence length, whether you use contractions, and how often you open a sentence with "I". Those three separate most people's natural writing from the default register of a cold email, and each is something you can look at and agree or disagree with. Nothing here tries to infer personality.

One good sample beats five. Paste the piece that sounds most like you at your best, not the longest thing you have.

8.6 Assets#

Assets in the sidebar, not under Settings, but it belongs in the same conversation.

Upload your CV, transcripts, statements, publication lists, project descriptions, writing samples, or notes. Tick Use this file when drafting on anything you want available as background when scoring matches and writing emails.

The table shows what you've uploaded, its type, whether it's used in drafting, and when it arrived. Your CV is the one that matters most — see chapter 1.

Getting a file back out. Each row has a download button, and the asset's own page has one too, plus a download link against every version in its history — the file you get is the original you uploaded, named after the asset's title. Nothing is ever only inside GradWorkbench.

Both tables have a search box, so a workspace with a document per professor stays usable.

Documents the platform writes#

The top half of the same page lists what has been generated from your profile. Ask your assistant for any of these by name:

Admin emails are written only when you ask for one explicitly. They are never produced as a follow-on step or as part of a batch — nothing should be drafting messages to a university's admissions office on your behalf unless you asked in that moment.

Everything else on the list is read by a professor or a committee, so it goes through your writing rules, and each document records the version it was written under.

Getting one out. A document's page has an Export card with four ways out: Markdown, LaTeX, plain text and PDF. The first three show you the rendered text before you take it, and download as .md, .tex and .txt. PDF opens the typeset document in your browser's viewer, from where you can save it — that is the one to attach to an application or send to someone. LaTeX remains the route to a document you want to restyle yourself: it is complete and compiles in Overleaf as-is.

Every format is rendered from the stored document at the moment you ask for it, so a typo you fixed a minute ago is fixed in all four. The PDF is typeset in Noto Sans, which covers Latin in full — every European accent, Vietnamese — along with Greek and Cyrillic, so names set properly rather than being flattened. Chinese, Japanese, Korean and Arabic are the exception: they come out as ?, and LaTeX is the route for a document that needs them.

What's next#

Questions and fixes →