Professor Analyzer

Professor / lab analyzer

This is the load-bearing skill of the PhD copilot: outreach emails, proposals, and ranking all consume the profile it produces. The goal is a profile so specific that a cold email built from it could only have been written to this professor — the opposite of generic.

What you produce

A single file at knowledge-base/professors/.md following the professor schema in shared/schemas/README.md (read it for the exact front-matter fields and section headings). Match that shape exactly so downstream skills can parse it.

Step 1 — Load context

Read knowledge-base/profile/profile.md (and profile/cv-master.md if fit detail is needed). You cannot assess "fit" without knowing the applicant's interests, background, and dealbreakers. Handle the applicant context in this order, so you never block unnecessarily:

1. If profile.md has real content, use it.
2. Otherwise, if the user's request itself describes their background and interests

(e.g. "I work on sim-to-real RL, analyze Prof X"), use that as the working profile — don't stop to ask for what you've already been told.

1. Only if you have neither, ask the user for their research interests and background, since

fit scoring is meaningless without them.

Identify the professor from the user's request: a name + institution, a homepage, a Scholar profile, or a lab URL.

Step 2 — Gather evidence

Follow the retrieval priority in shared/references/connectors.md — it defines which tool to reach for and how to fall back cleanly. Prioritize recency: a PhD starts in 1–2 years, so what matters is where the lab is heading, not its decade-old greatest hits. Ground every later claim in a real source.

Get the recent publications from a real index, not memory. Run the bundled helper to pull the professor's last ~3 years of work with abstracts:

python scripts/scholar_lookup.py "" --institution "" --since 

It queries OpenAlex (falling back to arXiv) and returns titles, venues, years, DOIs, and abstracts as JSON. If it prints a NOTICE that the API was unavailable or returns no works, fall back to WebSearch (always available) and the browser tools for Google Scholar — and note in the profile that structured-index enrichment was unavailable. Never fill the gap by guessing publications from memory; an invented paper in an outreach email is fatal.

Record where the evidence came from in the enrichment front-matter field: openalex or arxiv if the helper returned works, websearch if you fell back to search, none if you could retrieve no real publication evidence at all (in which case say so plainly and treat the whole profile as provisional — a profile with enrichment: none should not seed a confident outreach email).

Then gather, roughly in priority order:

• Recent publications (last ~3 years) — from the helper / search above. For each, capture

the takeaway: what problem, what approach, what's new. This is what makes hooks specific.

• Lab website / "join us" / "prospective students" page. Current projects, stated open

problems, whether they're recruiting, funding mentions — and any application process / contact rules (see Step 2a).

• Grants / funding signals. Active grants suggest funded positions. Mark unknown when

unclear — don't guess.

• Research trajectory. Read across the recent papers for the direction: what thread is

the lab pulling on, what will the next few papers likely be about?

If a source can't be fetched, note the gap rather than fabricating around it.

Step 2a — Capture the contact / application rules (downstream skills depend on this)

Actively look for how this professor wants to be approached. This is not optional polish: outreach-email and opportunity-ranker branch on it, and getting it wrong wastes the applicant's one shot. Record each rule in the typed front-matter field so downstream skills can parse it without reading prose, and add any nuance to ## Notes:

• email_policy — set do-not-email-admissions when the page/profile says not to email

about admissions (many top labs do) and routes applicants to the program; welcomes-inquiry when brief intros are invited; unknown if the page is silent. In ## Notes, capture the routing it implies ("apply to and name me") and any subject-line convention.

• admission_model — direct-to-advisor, program-committee, rotation, or unknown.
• accepting_students — yes / no / unknown for the applicant's target year; note the

year in ## Notes if stated.

Set these to unknown rather than guessing when the source is silent — a wrong welcomes-inquiry is worse than an honest unknown. When you find a hard rule (e.g. do-not-email), surface it in your summary too: it changes whether outreach is even the right move.

Step 3 — Assess fit

This is the part that makes the profile useful. Compare the professor's trajectory against the applicant's profile and be honest and specific:

• Overlaps — concrete intersections between their open problems and the applicant's

interests/skills. Name the paper and the matching part of the applicant's background.

• Gaps — where the applicant lacks relevant background, or the lab's direction diverges

from their goals. Real assessment includes the misses.

• fit_score (0–100) — your overall judgment, justified by the overlaps/gaps above, not

a vibe. Reserve 80+ for strong, well-evidenced matches.

• Funding & "accepting students" signals, set honestly to unknown when unclear.

Step 4 — Outreach hooks

The payoff section. List 2–4 specific things the applicant could reference in a first email: a particular recent paper and a genuine, substantive reaction or question; a connection between one of their projects and the applicant's work; an open problem the applicant is positioned to contribute to. These must be real and specific — they are what separates an authentic email from spam. No flattery, no invented enthusiasm.

If Step 2a found that the professor doesn't take cold admissions emails, these hooks still matter — note that they're best used in the statement of purpose / program application (or a narrow, substantive research question) rather than a "please take me" cold email, so the downstream skills route them correctly.

Step 5 — Write the file

Write knowledge-base/professors/.md with all schema front-matter fields filled (blank, not absent, when unknown) — including the typed funding_signal, accepting_students, admission_model, email_policy, and enrichment fields the downstream skills read — and a ## Sources section with one link per claim-bearing source. Then give the user a short summary: the fit verdict, the single strongest hook, and any gap they should be aware of. If you found a clear opening, suggest the natural next step (e.g. drafting outreach).

Guardrails

Follow shared/references/ethics.md. The cardinal rule: never invent a publication, finding, grant, or shared interest. If you're unsure whether something is real, mark it uncertain and cite what you actually found. A profile that honestly says "weak fit" is more valuable than a flattering one that wastes the applicant's outreach on a bad match.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: bydeng01
• Source: bydeng01/phd-application-skill
• License: MIT

Install and usage instructions live in the source repository linked above.