phdtaketaketake: Instructions file for Codex

phdtaketaketake — agent instructions

This file is the entry point that OpenAI Codex (and other agents that discover persistent instructions via AGENTS.md per the Codex convention) uses to find this skill. It is intentionally a short pointer, not a copy of SKILL.md. Maintaining two copies of the skill contract creates drift; the canonical contract lives in SKILL.md, and Cursor / Codex / any future host reads that file or gets a generated short-pointer like this one.

Use

When the user asks about PhD advisor matching, application triage, PI ranking, connection-first evaluation, evidence-backed application strategy, or PhD-application CV optimization, follow the workflow defined in SKILL.md:

• Workflow A (advisor matching): Steps 1–8.5 — gather profile, generate discovery plan, find candidates, compute connection edges, score with the 5-layer pipeline (CAPEG → application_strength → risk_adjusted → difficulty_adjusted → strategy bucket), present cards.
• Workflow B (CV optimization): Steps CV-1–CV-6 — read the bundled LaTeX template, fill from user-typed input, optionally reorder for a target PI from match.json, compile.
• Workflow C (admission probability): Steps AP-1–AP-4 — collect the required applicant inputs, research each target school / PI with evidence, run phdtaketaketake-admission, present probability + interval + levers.

Hard rules (non-negotiable for both workflows)

1. Connection-first — w_C > w_A in every tier. Verified academic network beats h-index.
2. Evidence-first — every claim traces to a real source the agent actually fetched. Four-state semantics: Verified / Verified-empty / Missing / Blocked. Strict mode rejects unsourced claims.
3. No invention — never fabricate candidate facts, advisor connections, publication records, GPA conversions, opportunity signals, CV experiences, papers, or skills.
4. CV source-of-truth — every fact in the rendered cv.tex traces to something the user explicitly provided in conversation. Target PIs from match.json drive ordering decisions; their names never appear in the CV body.
5. Probabilities only from the admission estimator. The matcher's output is a relative-fit index, not an admission probability. A numeric admit chance comes only from Workflow C (phdtaketaketake-admission), always with its 80% interval and never stated as a guarantee.