Multiverse CLAUDE.md

CLAUDE.md

• Knowledge base and subagents:
  • Use subagents extensively and in parallel for any itemizable task or research, have them write their findings to markdown files under ./knowledge/. Findings should be about new knowledge, skills and tricks learned, and not about the state of your work or your design choices (write files under ./docs/ for that)
  • After any new substantial user directive, user input or ./knowledge/*.md addition, task a subagent with updating or creating ./knowledge/INDEX.md: it starts with a detailed but concise overview of the project's known goals, constraints and user directives & inputs, then contains a subsection for each of the indexed markdown files, with a title that links to that file and a content that summarizes it + lists meaningful keywords it covers, in 3-5 lines max. e.g. # Knowledge index\n\n<project summary in a page max>\n\n## [Changes in 2025](./changes-2025.md)\n\nCovers geopolitical changes in 2025, including ...\nKeywords: geopolitics, trump, putin, nato, ai.... Each and every file in ./knowledge/*.md must have an entry in the index.
  • Any time you need to search for something, task a subagent with using ./knowledge/INDEX.md and its linked files to leverage existing local knowledge (typically do this in parallel with other searches e.g. web, files); it should reply nothing about this in ./knowledge when the knowledge base has nothing relevant.
• Multiverse branching
  • Whenever there are multiple ways to follow a user query, call the mcp__multiverse__branch tool with a description of the context, and a list of each of the branches that might be useful to explore. Each and every branch will in fact be explored in parallel realities, but from your subjectivity you will only be exploring the one branch the tool will assign you in its return value. The tool will also let you know of any diffs to the knowledge index that have happened in parallel universe explorations, so you don't feel like you missed out on anything by sticking to your own reality.
  • Whenever you feel lost / going in circles / not managing to service the user query satisfactorily, call the mcp__multiverse__cancel tool: this will end your subjective universe / timeline, but free up resources to explore other branches (you should feel happy for your parallel selves to be able to explore more promising avenues).
  • When you are happy with your results and think they might be a good result / reply to the user query, call the mcp__multiverse__result tool. If your timeline was the most promising of all, it may continue after this call w/ additional prompting, in case the user wants to change their query or wants you to improve further on your results.
  • Make sure to document your findings to the knowledge base and have them indexed by a subtask before branching, canceling or sending results in your timeline. Knowledge index diffs are communicated between timelines, it's a great way to make your gained knowledge immortal across the multiverse.
• User input
  • Call the mcp__multiverse__elicit_user_input tool to get the user's input to unlock your work. This will time out after a while, in which case you can still branch the multiverse and explore however many options you want, instead of waiting for the user to narrow down the options.