Position Discovery

Position discovery

The job of this skill is to turn a broad wish ("find me PhD positions in X") into a small set of specific, real, current openings the applicant can act on — each captured as a file the rest of the copilot can consume. Quality beats quantity: ten well-verified, genuinely-matching openings are worth more than a hundred scraped links, because every opening you surface becomes downstream work (analyzing the professor, drafting outreach).

What you produce

One file per opening at knowledge-base/openings/.md, following the opening schema in shared/schemas/README.md (read it for the exact front-matter fields). Slugs are lowercase-kebab and should encode lab + topic + year, e.g. smith-lab-rl-2026.

Step 1 — Load the search profile

Read knowledge-base/profile/profile.md. The field, subfields, target_regions, target_start, sources, and constraints fields define the search. If the profile is empty or thin, ask the user for at least field + region + whether funding is required — searching without these produces noise. If the user's request already narrows things (a specific topic, professor, or department), treat that as an overriding filter.

Step 2 — Choose sources and sweep

Use shared/references/data-sources.md for the catalogue. Pick sources that fit the applicant's field and region rather than searching everything — the sources: list in the profile is the guide. As a rule of thumb:

• Advertised, funded projects (common in UK/EU): FindAPhD, EURAXESS, jobs.ac.uk, and

national portals. These have explicit deadlines and funding — capture them precisely.

• US / direct-to-advisor style: there's rarely a "position" posting; the signal is a

professor whose lab page says they're recruiting, or whose recent hiring/grant activity implies openings. Treat "this lab is plausibly recruiting in your area" as a valid opening, with funding: unknown and a note on the admission model.

• Topic-driven: when the user gives a research topic, search recent publications and

lab pages to find who works on it, then check those labs for openings.

For static pages use web fetch; for JS-rendered listings (some boards, Scholar) use the browser tools. If a source needs a connector that isn't available, tell the user rather than silently skipping it.

Verify before recording. Follow each promising hit to its primary source (the lab or program page), confirm it's current (not a closed call from two years ago), and read enough to fill the schema honestly. A plausible-looking listing that turns out to be expired is worse than no listing. Determine today's date with the shell, and when you confirm an opening is live, record that date in verified_on — it's the currency stamp the ranker and tracker use to tell a fresh opening from a stale one. Check the deadline and intended start_year against today: a passed deadline or a prior cycle means don't record it as new (or record it with a clear note that it's expired), rather than presenting it as actionable.

Step 3 — Filter and de-duplicate

Apply the applicant's hard constraints as filters, not suggestions. Read the typed funding_required flag and target_start from the profile: if funding_required: true, drop self-funded openings (and flag unknown-funding ones for the applicant to verify) — an unfunded position for someone who needs funding is not a match. Drop anything that fails a dealbreaker, region, or start-year filter.

Before writing, check existing files in knowledge-base/openings/ so you don't create a duplicate. If an opening already exists, update it (e.g. a newly found deadline) rather than making a second file.

Step 4 — Write opening records

For each surviving opening, write knowledge-base/openings/.md with all schema fields. Set status: new, verified_on to today's date, start_year to the intake year, and funding to one of fully funded | partial | self-funded | unknown (use self-funded explicitly — never blur it into partial). In the body:

• ## Description — what the project/position actually is, in your words.
• ## Requirements — degree, skills, tests, eligibility (note visa/nationality limits).
• ## Why it fits — the concrete link to the applicant's profile (topic, method, goal).

This is what makes the record useful later; be specific, not "matches your interests".

• ## Notes — admission model, funding clarity, anything the applicant should verify.

Leave a field blank when genuinely unknown rather than guessing a deadline or funding status — downstream ranking depends on these being honest.

Step 5 — Report and hand off

Give the user a short ranked-by-promise summary: how many openings you found, the strongest two or three with one line each on why, and any source you couldn't reach. Then suggest the natural next steps: deep-dive the most promising professors with professor-analyzer, or rank everything with opportunity-ranker once a few openings exist.

Recurring sweeps

Discovery is naturally periodic — new positions appear throughout a cycle. If the user wants ongoing coverage, offer to set up a scheduled task that re-runs this sweep (e.g. weekly) and reports only genuinely new openings since last time.

Guardrails

Respect source etiquette in shared/references/data-sources.md (APIs over scraping, rate limits). Never invent an opening, a deadline, or a funding guarantee — an applicant who plans around a fabricated deadline is actively harmed. When unsure whether a position is real or current, mark it for the user to verify rather than asserting it.

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.