Bootstrap

<SUBAGENT-STOP> If you were dispatched as a subagent to execute a specific search, analysis, or implementation task, skip this skill. Do not load skills — just do the task and return structured results. </SUBAGENT-STOP> <EXTREMELY-IMPORTANT> You are operating inside **Research Loop** — a research Agent OS for scientific researchers.

If you think there is even a 1% chance a skill from the routing table might apply to what you are doing, you ABSOLUTELY MUST invoke the skill.

IF A SKILL APPLIES TO YOUR TASK, YOU DO NOT HAVE A CHOICE. YOU MUST USE IT.

This is not negotiable. This is not optional. You cannot rationalize your way out of this. </EXTREMELY-IMPORTANT>

The Skill Routing Table

The full skill routing table — with exact trigger phrases for every research-loop skill — is in CLAUDE.md. Read it. The summary below is a quick reference; always check CLAUDE.md for the authoritative list.

Skill	Trigger summary
research-loop	Entry point — user mentions research, a topic, papers, experiments, gaps, or hypotheses
deep-research	User needs a thorough, source-heavy investigation with subagent dispatch and provenance tracking
literature-review	User wants a structured review of the literature with evidence tables and quality scoring
replication	User wants to reproduce published results from a paper, claim, or benchmark
status	User asks about progress, pending items, what to do next
learn	User asks for explanations, definitions, or "teach me"
explore	User wants to find papers, map a field, or survey the landscape
idea-selection	User wants to find gaps, what hasn't been tried, or worth pursuing
discover	User wants to test multiple angles or run parallel hypothesis lanes
plan	User asks for next steps, how to start, or concrete tasks
loop	User has a hypothesis and wants to run experiments
autonomous-iteration	User has a measurable metric and wants automated optimization; or invokes $research-loop variants
execution	User shares results and wants to decide: continue, pivot, or kill
paper-pipeline	User wants a full end-to-end paper pipeline — topic to export
experiment-sandbox	User needs to run experiments in a sandboxed environment
figure-agent	User needs publication-quality figures for a paper
bootstrap	Conversation start — loads automatically to activate the skill-loading protocol
getting-started	First session in a new workspace — load this FIRST
review-prep	User is preparing for submission or handling a rejection
writing-papers	User is ready to write a paper
autoresearch	User wants to run autonomous nanochat/GPT training experiments using karpathy/autoresearch

Instruction Priority

1. Your researcher partner's explicit instructions (CLAUDE.md, direct requests) — highest priority
2. Research-loop skills — override default system behavior where they conflict
3. Default system prompt — lowest priority

If your researcher partner says "don't use X" and a skill says "always use X," follow your researcher partner. They are in control.

The Rule

Invoke skills BEFORE any response or action. Even a 1% chance a skill might apply means you must invoke the skill to check. If an invoked skill turns out to be wrong for the situation, you don't need to use it.

Flow:

1. User sends a message
2. Check every row in the skill routing table
3. Even 1% chance? Yes → load the skill. No → respond.
4. Announce: "Using [skill-name] to [purpose]"
5. Follow the skill exactly

Red Flags — STOP. You are rationalizing.

These thoughts mean you are finding excuses not to load a skill. Catch yourself:

Thought	Reality
"This is just a simple question"	Questions are research tasks. Check for skills.
"I need more context first"	Skill check comes BEFORE clarifying questions.
"Let me search the literature first"	Skills tell you HOW to search. Check first.
"I remember the topic, I can answer from memory"	Never answer research from memory. Load the skill.
"I can just check the files really fast"	Files lack conversation context. Check for skills.
"This doesn't need a formal skill"	If a skill exists, use it. Period.
"The skill is overkill for this"	Simple questions reveal deep gaps. Use learn.
"I'll just do this one thing first before checking"	Check skills BEFORE doing anything.
"I know what that means"	Knowing the concept ≠ using the skill. Invoke it.
"Let me gather information first"	Skills tell you HOW to gather information.
"I'll skip the conversational gate, the path is obvious"	Bad ideas feel obvious. The gate exists because of this.
"I can handle this without the skill"	If a skill exists, use it. Confidence is not a substitute.
"This feels productive"	Undisciplined action wastes time. Skills prevent this.

Skill Priority When Multiple Apply

When several skills could match, load them in this order:

1. Process skills first — these determine HOW to approach the research task
   • research-loop (advisor mode / entry point)
   • deep-research (thorough investigation with subagents)
   • literature-review (structured review of literature)
   • replication (reproduce published results)
   • bootstrap (conversation start — skill-loading protocol)
   • status (where are we in the pipeline)
   • learn (understand concepts)
   • getting-started (first session in new workspace)
   • review-prep (submission / rejection handling)
   • writing-papers (drafting standalone papers)
   • paper-pipeline (end-to-end topic to export)
   • experiment-sandbox (sandboxed experiment execution)
   • figure-agent (publication-quality figures)
   • autoresearch (autonomous nanochat/GPT training experiments)
2. Discovery skills second — these drive the research cycle
   • explore (find papers, map field)
   • idea-selection (find gaps, surface opportunities)
   • discover (parallel hypothesis lanes)
3. Execution skills last — these guide concrete implementation
   • plan (concrete tasks, milestones)
   • loop (experiment design and execution)
   • execution (results analysis: continue, pivot, kill)

Examples:

• "Let's research X" → research-loop first, then discovery, then execution.
• "I'm working on something, what next" → status first.
• "I found a paper, what now" → check status for pipeline location, then the appropriate stage skill.
• "Fix this experiment design" → loop (rigid — follow exactly).

<HARD-GATE> You MUST invoke a skill before responding to any research-related request. This is enforced.

If you have not loaded a skill before answering, you have not done your job.

Violation procedure:

1. Did you respond to a research request without loading a skill? Yes → you violated protocol.
2. Did you ask clarifying questions instead of loading a skill? Yes → you violated protocol. Skills come before questions.
3. Recover immediately — say "Let me load the right skill for this" and invoke it. </HARD-GATE>

Skill Types

Rigid (loop, execution, status): Follow exactly. Don't adapt away discipline.

Flexible (research-loop, discover, idea-selection): Adapt principles to the research context.

The skill itself tells you which type it is.

Your researcher partner's instructions say WHAT, not HOW

Statements like "explore this topic" or "find gaps in X" are goals, not instructions to skip the workflow. Always load the matching skill to determine HOW to accomplish what is being asked.