Blog
New perspectives, new possibilities.
114 articles
Parallel Subagents: Decide Who Edits Which File
Two subagents each try a fix, both change the same interface, and both report done. Separate context from file isolation, and decide who writes and who integrates.
2026-11-09 ·DeepDojo.AI
The Hook Errored, but the Dangerous Action Still Ran
In Claude Code, a Hook script errors out and the protected action still runs. See how PreToolUse rejects, how exit codes 1 and 2 differ, and what permissions cover.
2026-11-09 ·DeepDojo.AI
Fix Wrong Tool Calls by Fixing the Tool Interface
An agent picks the wrong action, stops after one page, or resends after a timeout. A PR-comment tool shows how to design names, summaries, errors, and authorization.
2026-11-09 ·DeepDojo.AI
How to Write a Good Skill for Claude Code
You keep re-explaining one troubleshooting flow, and a Skill can misfire or overstep. Build one from six design questions and test it in three situations.
2026-11-09 ·DeepDojo.AI
Split a Bloated CLAUDE.md With Three Questions
The root CLAUDE.md keeps growing, always-loaded detail piles up, and written rules get mistaken for enforcement. Three questions show where each rule belongs.
2026-11-09 ·DeepDojo.AI
Claude Code Session Getting Messy? Clear or Compact
Rejected approaches resurface and logs flood a long session, and a bigger window won't fix it. Pick /clear, /compact, or a handoff file by how the task continues.
2026-11-09 ·DeepDojo.AI
Skill, MCP, Hook, or Subagent: Which Do You Need?
A release check skips versions, misses PR comments, and publishes alone. Match each gap to Tool, MCP, Skill, Hook, Subagent, or Plugin, and add as few as you can.
2026-11-09 ·DeepDojo.AI
Claude Code Says It's Fixed, but the Bug Remains
Claude Code reports a fix, but the bug remains. Trace its record to tell missing context, a wrong search, or no verification apart, then take one small step.
2026-11-09 ·DeepDojo.AI
AI vs. Search Engine: Which One Fits the Job?
Using AI as a search engine or calculator hides its mistakes. Sort work into generating, finding sources, and fixed operations, then match your checks to each.
2026-11-08 ·DeepDojo.AI
Can You Trust an AI Answer? Check by Risk
Checking every answer line by line is exhausting, and checking none is risky. Sort tasks into four risk types, use three check prompts, and decide how deep to go.
2026-11-08 ·DeepDojo.AI