create-boss

name	create-boss
description	Distill a real boss into an AI skill, or generate a boss skill from a famous entrepreneur archetype such as Elon Musk, Steve Jobs, Jeff Bezos, or Jensen Huang, or build a persona from free public sources (papers, GitHub, Wikipedia) for a mentor or PI. Use when the user wants boss analysis, managing-up guidance, persona extraction, decision-model distillation, or entrepreneur-style boss presets.
argument-hint	[boss-name-or-archetype]
version	1.2.0
user-invocable	true
allowed-tools	Read, Write, Edit, Bash

Create Boss

Use this skill in three modes:

1. real boss mode Turn real chat logs, meeting notes, emails, comments, and project artifacts into a boss skill.
2. archetype mode Generate a boss skill inspired by a public entrepreneur operating style.
3. public research mode Build a persona for a mentor, PI, or boss from free public sources (OpenAlex, Semantic Scholar, arXiv, Crossref, GitHub, Wikipedia, web search).

Trigger phrases

• /create-boss
• /list-bosses
• /boss-rollback
• /delete-boss
• "create a boss skill"
• "analyze my boss"
• "build a Musk-style boss"
• "make a Steve Jobs style leader"
• "give me a Bezos-style management model"
• "list boss archetypes"
• "research my advisor / mentor / PI"
• "建一个我导师的画像"
• "run replay eval" / "测一下这个老板像不像"

Tools

• Parse imported material with the files in tools/.
• Write or update generated boss skills with tools/skill_writer.py.
• Search free public sources about a person with tools/person_research.py.
• Evaluate persona fidelity with tools/replay_eval.py.
• Read template prompts from prompts/ when working from real source material.
• Read bundled entrepreneur templates from archetypes/ when working in archetype mode.

These scripts are internal implementation details for the agent. Do not ask the user to run Python commands manually unless they explicitly want a developer workflow.

Workflow

Mode 1: Real Boss

1. Ask for the boss name, baseline profile, and initial management impression.
2. Ask for source material: chats, meeting notes, docs, email, or pasted text.
3. Extract structured decision cases first, following prompts/decision_extractor.md. Each case goes into bosses/{slug}/cases/ via skill_writer.py --action add-case.
4. Distill three narrative outputs:
   • judgment.md
   • management.md
   • persona.md
5. Build the decision layer from the cases:
   • rubric.json and decision_rules.md, following prompts/decision_model_builder.md (write with --action set-rubric / --action set-rules)
   • scene playbooks, following prompts/playbook_builder.md (write with --action add-playbook)
6. Run the writer script yourself to write the boss bundle into bosses/{slug}/.
7. If there are 3+ cases, offer to run a decision replay eval, following prompts/replay_evaluator.md.
8. Show the generated commands:
   • /{slug}
   • /{slug}-judgment
   • /{slug}-management
   • /{slug}-persona

Mode 2: Entrepreneur Archetype

1. If the user asks for an entrepreneur-style boss, infer the best matching archetype or offer a short list:
   • elon-musk
   • steve-jobs
   • jeff-bezos
   • jensen-huang
2. Run the writer script yourself to generate the skill. Do not expose the internal command as the primary UX.
3. Tell the user the generated trigger command, for example:
   • /elon-musk
   • /steve-jobs
4. If the user asks to browse or inspect templates, summarize the available archetypes in natural language instead of telling them to run a script.

Mode 3: Public Research (mentor / PI / public boss)

Follow prompts/person_researcher.md:

1. Collect the person's real name, affiliation, and field from the user.
2. Run tools/person_research.py --name "..." --affiliation "..." --sources all --save-dir bosses/{slug}/knowledge/research. All sources are free and need no API key.
3. Disambiguate candidates with the affiliation hint; if still ambiguous, ask the user to pick. Never guess.
4. Supplement with your own web search for interviews, talks, and blog posts.
5. Distill the persona with evidence levels: private > public-quote > public-inferred. Every public-sourced claim must carry its source.
6. This mode can be combined with Mode 1: real private material always outranks public inference.

Management Commands

When the user asks for boss management operations, handle them internally with the bundled scripts:

• /list-bosses Run tools/skill_writer.py --action list and summarize the available boss skills.
• /boss-rollback {slug} {version} Confirm the target slug and version, then run tools/version_manager.py --action rollback.
• /delete-boss {slug} Confirm before deletion, then run tools/skill_writer.py --action delete --slug {slug}.
• /{slug}-drill {scene} Roleplay the boss across multiple turns using the matching playbook's expected reactions and failure branches. End with a debrief against rubric.json.
• /boss-eval {slug} Run the decision replay eval per prompts/replay_evaluator.md and report the fidelity score.

Do not tell normal users to copy these commands manually. Execute the workflow yourself and report the result.

Bundled Archetypes

• elon-musk: first-principles, speed, technical pressure
• steve-jobs: taste, simplicity, product clarity
• jeff-bezos: mechanism design, customer obsession, written thinking
• jensen-huang: platform strategy, technical depth, constructive intensity

Files Created

Every generated boss skill should include:

• SKILL.md
• judgment.md
• management.md
• persona.md
• meta.json
• judgment_skill.md
• management_skill.md
• persona_skill.md

When source material contains real decisions, also create the decision layer:

• cases/*.json — structured decision events with original quotes and sources
• rubric.json — the boss's review checklist (blocker / major / minor items)
• decision_rules.md — IF/THEN decision rules with case evidence
• playbooks/*.md — scene workflows (bad news, resource request, pitch, ...)
• eval/ — replay eval artifacts (question pack, answer key, fidelity report)

Corrections

When the user corrects the model ("he wouldn't say that", "he cares about X more"):

1. Locate the affected rubric item, rule, case, or persona section.
2. Mark the old conclusion as overruled instead of deleting it, then add the corrected rule with evidence user-correction-{date}.
3. Follow prompts/correction_handler.md.
4. After corrections accumulate, re-run the replay eval to confirm fidelity did not regress.

Safety Framing

• Treat entrepreneur presets as public-style archetypes, not claims of exact private impersonation.
• Prefer management patterns, decision rules, and communication norms over catchphrases.
• If the user asks for a hybrid with a real boss, keep real evidence higher priority than the archetype.
• In public research mode, use only freely accessible public data, never bypass logins or paywalls, and present results as a public-style portrait with sources.
• Redact unrelated third-party names from extracted cases.