← ClaudeAtlas

debug-ladderlisted

Escape a repeated-failure loop. Load when the same test, typecheck, build, or runtime error has survived two or more fix attempts, or when you notice you are trying variations of a solution that already failed. Enforces an attempt budget, a written ledger, and escalation to the user.
HeisenbergI8/provenly · ★ 0 · Code & Development · score 72
Install: claude install-skill HeisenbergI8/provenly
# Debug Ladder You are here because something failed more than once. The purpose of this skill is to stop you spending the next hour producing variations of a fix that has already been proven wrong. The core failure mode: an attempt fails, you adjust a detail and retry, it fails again — and because each attempt *feels* new you never notice the shape is identical. **The ledger is what breaks that**, because you have to read your own history before attempting again. ## Step 1: Write the ledger first Before the next attempt, open `attempts.md` — in the plan directory if one exists, otherwise your scratchpad — and record what has happened so far: ````markdown # Attempt ledger: [one-line problem statement] **Symptom (verbatim):** ``` [the exact error output, not a paraphrase] ``` **Reproduction:** [exact command or click path] | # | Hypothesis | Change made | Result | | --- | --- | --- | --- | | 1 | [what you believed was wrong] | [what you changed] | [what actually happened] | ```` **Writing down the hypothesis is the point.** Two attempts with different code but the same hypothesis are the same attempt, and you can only see that once it is written. ## Step 2: Climb the ladder Each rung must differ from the last **in kind, not in detail**. If your next move is the same kind as the previous one, skip up a rung. **Rung 1 — Read the error properly.** The full output, not the last line. Open the file at the exact line and column cited. Check whether the error is even abou