update-knowledge

Update Knowledge Base

NON-NEGOTIABLE: ALL PATHS ARE RELATIVE TO THE REPO ROOT All .claude/planning-with-james/ paths in this skill are relative to the current working directory (the repo you're working in), NOT ~/.claude/. The knowledge graph lives inside the project, not in your home directory. If you're unsure, run pwd to confirm you're in the repo root.

NON-NEGOTIABLE: PROGRESS TRACKING After every module update, you MUST update _update_progress.json BEFORE starting the next module. This file is your lifeline through context loss. If you skip this step and autocompact hits, all work since the last progress write is lost.

NON-NEGOTIABLE: LEAN MAIN CONTEXT The orchestrator (you) should hold minimal state. Do NOT read all module _index.md files, _graph.json, or _architecture.md into your context during standard updates. That is wasted context. Subagents read their own knowledge. You read only what you need to plan.

NON-NEGOTIABLE: LINEAR EXECUTION Process modules ONE AT A TIME. Do NOT launch parallel subagents for module updates. Each module update should: spawn agent -> receive brief summary -> update progress -> next. This prevents context exhaustion.

NON-NEGOTIABLE: AGENT PERMISSIONS ALL agents you spawn (Task tool) MUST include mode: "bypassPermissions". Background agents that hit permission prompts will have their Write/Edit calls silently denied — they burn through turns doing research but never write output files. The orchestrator then polls for result files that will never exist. This destroys all agent work. Always include this parameter on every agent spawn, background or foreground.

NON-NEGOTIABLE: NO BASH FOR FILE MODIFICATIONS Do NOT use Bash to modify knowledge files. No sed, awk, for loops, echo redirection, or cat heredocs for writing or editing files. ALWAYS use the Read → Edit or Read → Write tools for file changes. Bash is ONLY permitted for: git commands (git diff, git rev-parse, git log) and running project tools (linters, test runners). This applies to you (the orchestrator) AND to all subagents you spawn — include this rule in every subagent prompt. Violating this blocks unattended operation with permission prompts that require manual approval.

---