← ClaudeAtlas

log-mistakelisted

Write a logbook entry when something breaks, works unusually well, or a hard-to-reverse choice gets made. Use the moment it happens, not at the end of the session.
sheldor26/Blastdoor · ★ 0 · AI & Automation · score 67
Install: claude install-skill sheldor26/Blastdoor
# Log it now The entry gets written when the thing happens, not at the end of the session. At the end the details are gone, and it is the first task to get cut when the session is running long. Four seconds now, or nothing ever. ```bash node .bitacora/cli.mjs new mistake "Specific title" --tags area,failure-mode --severity high node .bitacora/cli.mjs new learning "Specific title" --tags area node .bitacora/cli.mjs new decision "Specific title" --tags area ``` The command assigns the id and scaffolds the sections. You write the prose. ## If a subagent found it, log it in the main session A subagent runs in its own context window. It follows the main `CLAUDE.md`, so it can read this skill and run `recall` — but when it returns, it returns a summary, and everything specific it saw is gone with its context. That matters more than it sounds, because subagents get the exploratory and verification work: reading unfamiliar code, checking whether something is actually broken. That is exactly where mistakes are discovered. So when a subagent reports something that broke or surprised it, the main session writes the entry, and writes it before moving on — the detail is in the summary you are holding right now, and nowhere else. If the summary is too thin to support a real **Guardrail**, ask the subagent for the specifics before its findings go stale, rather than filing a vague entry. ## Which kind - **Mistake** — anything broke, surprised us, or had to be fixed twice. A near-