← ClaudeAtlas

reprolisted

Find where a Claude Code session went wrong, reproduce it in a sandbox, and fix the prompt. Tells a model that ignored its instructions from a missing or vague instruction, or a vague operator. Proposes a text fix, proves it with reruns, and changes nothing until the user says yes.
xtompie/repro · ★ 0 · AI & Automation · score 75
Install: claude install-skill xtompie/repro
Talk in the user's language. Ask one question at a time. Never edit the user's files before step 8. Run every shell command on its own: no `;`, `&&`, `|`, no variables like `S=...` or `$S`, no `cd`. Write paths out in full. Plain permission mode rejects combined commands, and the user then has to approve each one. ## 1. Read the session - No session given: this session, id `${CLAUDE_SESSION_ID}`. - An argument that ends in `.jsonl` or contains `/` is a file path: use it. Anything else is a session id: find the file with `ls ~/.claude/projects/*/<id>.jsonl`. Several: take them one by one. - Convert it: `perl ${CLAUDE_SKILL_DIR}/scripts/session-text.pl <file>`. Read the output, not the JSONL. Turn numbers like `[12]` are how you point at moments. - Note which instructions were in play: the skills that ran, `AGENTS.md`, `CLAUDE.md`, docs the model read. Read them. ## 2. Find suspects If the user said what went wrong, go to that moment. Otherwise list moments where the result differed from what the user wanted: corrections, interrupts, rejected tool calls, "done" followed by a failure, an instruction broken in plain sight. For each suspect pick one cause. Read `${CLAUDE_SKILL_DIR}/references/causes.md` first. - **ignored**: a clear instruction was loaded and the model did otherwise. - **gap**: the instruction was missing or imprecise; the model did a reasonable thing with what it had. - **operator**: the user was vague, contradictory, or expected something never said. - **