← ClaudeAtlas

forge-debuglisted

This skill should be used when implementation is stuck or the agent keeps getting something wrong in a Context Forge project — phrases like "forge-debug", "this keeps failing", "the agent can't fix this", "we're going in circles", "stuck on a bug", or "it broke again". It runs a disciplined stop-and-diagnose strategy instead of thrashing with more guesses.
yerros/context-forge · ★ 0 · AI & Automation · score 75
Install: claude install-skill yerros/context-forge
# forge-debug When the same problem resists two or more fix attempts, stop guessing. More attempts without a model of the failure just compounds damage. Switch to this disciplined loop. (A fresh bug report that isn't stuck yet enters through `forge-fix`, which triages and hands off here when the cause is non-obvious. After this loop resolves it, closing — tracker, lesson, `fix/` branch — follows `forge-fix` step 7.) ## Argument Text after the command describes the stuck problem (e.g. `/forge-debug the websocket reconnect keeps failing`) — treat it as the starting symptom for step 1. No argument → take the failure from the current conversation, or ask for it precisely. ## Stop condition If you've tried to fix something twice and it still fails, do NOT try a third blind fix. Run this process. ## The loop ### 1. State the failure precisely Write down: what was expected, what actually happens, the exact error or wrong output, and when it started (which unit/change introduced it). Vague symptoms produce vague fixes. **Read the attempt log first** (the `attempt N:` lines under the unit's In Progress entry in the tracker, per the loop contract): every approach listed there is already known to not work — diagnosis starts from that knowledge, and no hypothesis below may be a relabeled version of a logged failed attempt. If the project has a retrieval index, also query it for prior encounters — `bash "${CLAUDE_PLUGIN_ROOT}/skills/forge-init/scripts/forge-index.sh" query "<sym