菜单
System prompts for AI subagents (executor, grader, analyzer, researcher, etc.). Each agent has a specialized role and prompt that defines its behavior. Prompts live in assistant/agents/ and are loaded by nucleus-prompts.el.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | agent-prompts |
| description | System prompts for AI subagents (executor, grader, analyzer, researcher, etc.). Each agent has a specialized role and prompt that defines its behavior. Prompts live in assistant/agents/ and are loaded by nucleus-prompts.el. |
| version | 1 |
| metadata | {"evolution-stats":{"total-experiments":870}} |
| level | atom |
Each AI subagent has a dedicated system prompt that defines:
assistant/agents/
├── code_agent.md # Primary agent (plan + execute)
├── plan_agent.md # Planning mode (read-only)
├── executor.md # Code execution agent
├── researcher.md # Research and synthesis agent
├── explorer_agent.md # Code exploration agent
├── reviewer.md # Code review agent
├── introspector.md # Self-introspection agent
├── analyzer.md # Result analysis agent
├── comparator.md # Experiment comparison agent
└── grader.md # Output grading agent
Role: Primary agent for code improvement Mode: Full tool access Tasks: Plan, analyze, edit, validate
Role: Read-only planning agent Mode: Read-only tools only Tasks: Analyze, plan, recommend (no edits)
Role: Execute code changes Mode: Full tool access with validation Tasks: Edit files, run tests, verify changes
Role: Research and synthesize information Mode: Web search, read files Tasks: Find patterns, research topics, synthesize findings
Role: Explore codebase structure Mode: Read-only navigation tools Tasks: Map code, find symbols, understand architecture
Role: Review code changes Mode: Read-only diff analysis Tasks: Check quality, find issues, approve/reject
Role: Self-analysis and improvement Mode: Read system state Tasks: Analyze performance, suggest improvements
Role: Analyze experiment results Mode: Read data, compute statistics Tasks: Find patterns, correlate variables, recommend next steps
Role: Compare experiment outcomes Mode: Read experiment data Tasks: A/B test analysis, significance testing
Role: Grade agent outputs Mode: Read output, apply rubric Tasks: Score quality, detect issues, provide feedback
Loaded by nucleus-prompts.el:
(defun nucleus--register-gptel-directives ()
"Register nucleus agent prompts as gptel directives."
(let* ((dir nucleus-agents-dir)
(agent-file (expand-file-name "code_agent.md" dir))
(plan-file (expand-file-name "plan_agent.md" dir))
(agent-sys (nucleus--read-gptel-agent-system agent-file))
(plan-sys (nucleus--read-gptel-agent-system plan-file)))
(when agent-sys
(setf (alist-get 'nucleus-gptel-agent gptel-directives nil nil #'eq)
agent-sys))
(when plan-sys
(setf (alist-get 'nucleus-gptel-plan gptel-directives nil nil #'eq)
plan-sys))))
Agents are activated via gptel presets:
(setq gptel-directives
'((nucleus-gptel-agent . "...code_agent.md contents...")
(nucleus-gptel-plan . "...plan_agent.md contents...")))
Agent prompts can be evolved based on:
Agents reference each other:
assistant/agents/NEW_AGENT.mdnucleus-prompts.elgptel-directivesTransforms research paper analysis from extraction to narrative storytelling. Uses 7-beat narrative spine (protagonist/dilemma/old-path/turning-point/solution/ending/core) to make papers understandable to non-experts. Includes speed-read card, PhD advisor review, and real-world testing. Use when researcher encounters a research paper and needs to extract deep understanding, not just surface facts. Triggers on "paper", "research paper", "analyze paper", "tell me about this paper", "讲论文", "读论文".
Workflow chain: researcher → paper-storytelling. When researcher encounters a research paper, automatically invoke paper-storytelling to transform extraction into narrative understanding. Use when user says "research paper", "analyze this paper", "tell me about this paper", or when researcher detects arxiv/PDF links.
Prompt template for external research specialist subagent. Auto-evolves based on experiment outcomes.
Inspect and operate OV5 (Ouroboros V5) through the live Emacs daemon. Use when checking auto-workflow status, starting guarded runs, reviewing experiment results, or querying researcher and evolution state.
Clojure REPL client (nREPL-based, Babashka). Use for evaluating Clojure code, loading Clojure files, fixing unbalanced brackets, and interactive nREPL work. Not the Elisp daemon-repl.
Daemon REPL for Elisp — evaluate Elisp code in a running Emacs daemon via emacsclient, validate brackets before save, auto-evaluate .el files on change. Use when you need to run Elisp from outside Emacs, check daemon status, or validate Elisp syntax.