Advisor Profile Maintainer

Maintain a private, evidence-backed model of how a specific advisor thinks, communicates, evaluates work, and interacts with the user. This skill is for durable local profile maintenance, not public professor scouting and not impersonation.

Core Principles

• Treat user-provided private material as sensitive by default.
• Separate evidence, interpretation, and advice.
• Prefer small confirmed profile updates over large speculative rewrites.
• Store profiles as local files only when the user asks or approves.
• Never silently overwrite existing profile content.
• Never present the profile as the advisor's actual intent; describe it as a working model based on available evidence.

Profile Modes

1. Create a new advisor profile.

   • Use when the user wants to start tracking a mentor, advisor, PI, committee member, or collaborator.
   • Ask for the profile name/slug only if not inferable.
   • If the user provides raw or summarized material, extract evidence first, then propose a starter profile.

2. Update an existing profile.

   • Use when the user provides new meeting notes, email/chat excerpts, comments, feedback, oral remarks, or observations.
   • Read the existing profile files if paths are known.
   • Produce a proposed update diff and ask for confirmation before writing.

3. Consult a profile.

   • Use when the user asks how this advisor might react, how to prepare for a meeting, how to revise a draft for them, or how to interpret feedback.
   • Cite profile evidence and confidence levels.
   • Do not invent new profile facts during consultation unless the user asks to update.

4. Audit or consolidate a profile.

   • Use when the profile is stale, contradictory, too speculative, too long, or mixed with raw private logs.
   • Preserve raw evidence references; demote unsupported claims; mark conflicts.

Default Local Layout

When the user asks to persist a profile and does not specify a path, propose a repo-local or user-private location before writing:

advisor-profiles/
  <advisor-slug>/
    profile.md
    preferences.md
    interaction-log.md
    evidence.jsonl
    update-log.md

Use references/profile-schema.md for file contents and section names.

Workflow

1. Establish profile target and storage boundary.

   • Identify the advisor/person and relationship type.
   • Determine whether this is a create, update, consult, or audit request.
   • If writing files, confirm the target directory unless the user gave one.
   • Do not store raw private messages unless the user explicitly asks; prefer summarized evidence snippets.

2. Intake material.

   • Accept meeting notes, message excerpts, email summaries, paper comments, review comments, draft annotations, user observations, or public links.
   • Read references/private-source-handling.md before processing private or sensitive material.
   • If the user only gives broad impressions, mark them as user impressions rather than evidence.

3. Extract evidence units.

   • Convert material into small entries: date if known, source type, short excerpt or paraphrase, topic, observed behavior, possible interpretation, confidence, sensitivity.
   • Use evidence levels:
     • A: direct repeated evidence from original messages/comments or multiple meetings.
     • B: direct single evidence from one original note/comment.
     • C: user summary or recollection.
     • D: inference, hypothesis, or pattern needing confirmation.
   • Keep contradictory evidence instead of smoothing it away.

4. Map evidence to profile dimensions.

   • Research taste and intellectual standards.
   • Writing and presentation preferences.
   • Experimental or methodological standards.
   • Feedback and criticism style.
   • Communication habits and timing.
   • Decision patterns and risk tolerance.
   • Mentoring boundaries and support style.
   • User-specific relationship context.
   • Known red lines, sensitivities, or recurring misunderstandings.

5. Propose updates before writing.

   • Show a concise "candidate profile update" with additions, revisions, downgraded claims, conflicts, and open questions.
   • Label each update with evidence level and source reference.
   • Ask for confirmation before editing persisted files.

6. Write or update files only after approval.

   • Append evidence entries to evidence.jsonl.
   • Update synthesized claims in profile.md and preferences.md.
   • Add a dated note to update-log.md.
   • Append only relevant summarized events to interaction-log.md.
   • Avoid duplicating the same fact across files unless one entry is raw evidence and another is synthesis.

7. Use profile for advice.

   • When helping draft emails, meeting agendas, revision plans, or response strategies, cite the profile signals used.
   • Separate "likely advisor concern" from "recommended user action".
   • If confidence is low, say what new evidence would improve the model.

External Search

• Default to no external search.
• Use $anysearch only when the user explicitly asks to enrich the profile with public facts or current context, such as a new paper, lab page update, public talk, or venue preference.
• Never paste private notes, emails, chat logs, personal identifiers, sensitive relationship details, or unpublished work into external search.
• If external context is needed, search anonymized or public-only terms.

Output Shapes

For profile creation or update:

• Profile target
• Material processed
• Evidence units extracted
• Candidate profile updates
• Conflicts or uncertainty
• Proposed file changes
• Confirmation needed before write

For consultation:

• Question
• Relevant profile signals
• Likely advisor concerns
• Recommended strategy
• What to say / avoid
• Confidence and missing evidence

Boundaries

• Do not impersonate the advisor as if they are present or consenting.
• Do not generate manipulative psychological tactics.
• Do not store secrets, medical details, family matters, protected traits, raw credentials, or full private chat logs unless the user explicitly requests and the storage path is private.
• Do not infer personality, ethics, or intentions from a single weak signal.
• Do not let user frustration become a one-sided negative profile; preserve counter-evidence and benign explanations.
• Do not use this skill for public professor discovery; route that to $professor-fit-analyzer.

Source-Aware Notes

This skill is inspired by public mentor-distillation, person-profile, relationship-memory, and context-engine skills. Adapt those ideas conservatively: in academic advising, the goal is preparation and relationship memory, not simulating a real person's authority.