kenjaku: Skill for Claude Code

Sync sources — Fan-out/fan-in architecture (internal reference)

This is not a user command. This file documents the sub-agent architecture of Phase 2 of the main flow (sync sources in background — see CLAUDE.md). You never trigger the pull by hand: it's the question that triggers it. Note: /sync is a separate command that synchronizes the git repo between machines.

🔧 Adapt to your connectors. The examples below reference MCP tools generically (mcp__<slack>__…, mcp__<drive>__…, mcp__<calendar>__…). Replace them with the real names of the connectors you've wired up (see SETUP §6). Without a wired connector, this skill does nothing — the RAG engine answers on its own from the vault.

Absolute constraint

READ-ONLY. Never send a message, an email, or a reaction. Never post in a channel. Produce only local markdown files in the vault.

Sub-agent tooling — NEVER a shell to process text

The sub-agents are LLMs: they read and summarize by reasoning, not via the shell. Forbidden to use python3 -c, python, node -e, awk, sed, jq, grep, cat, head, tail — or any Bash command — to parse, load, split, slice or summarize content. Why this is non-negotiable (especially on Claude Desktop, Code tab):

• each ad-hoc command is unique → it re-triggers an authorization prompt on every call (endless prompts, impossible to pre-authorize);
• some (multi-line, # inside an argument, redirections) are refused for security and don't even offer "Always allow" — the user cannot accept them.

Instead:

• Read content → the Read tool (a vault file; or a large tool result that Claude offloaded to …/tool-results/…: read it with Read, not with python3 -c "open(...)").
• Write the raw source / the briefing → the Write / Edit tools.
• Split ("up to the Details section", "the first 4000 characters"…) → in your head, not in Python.