IDCLIP — No-Clip 👻
Doom: IDCLIP = you walk through walls. Here: if you're stuck behind a wall, don't headbutt it again — find a way around. The point is breaking the failure loop.
Most stalls last not because the problem is hard, but because we keep varying the same failing step. No-clip steps out of that: back off, name the wall, find a different route.
The protocol
1. 👁️ Name the wall — precisely
What's blocking, exactly? The error message, the missing dependency, the permission, the misunderstood API? If the wall is vague, illuminate it first (read the full error, the stack trace) — you can't go around it blind.
2. 🚪 List at least 3 routes around
Don't fixate on the first idea. Surface several materially different routes, e.g.:
- a different tool / library / approach to the same goal,
- sidestepping the problem (different order, different layer, a temporary stub),
- questioning the assumption ("do we even need this step at all?").
3. 🎯 Try the most promising
Pick the best odds/cost one and do it. If it also hits a wall, don't repeat — back to the list for the next.
4. 🧱 Don't hit the same wall twice
The mode's golden rule. If a route fails the same way as before, it's no longer a candidate. Note what didn't work and move on.
🦷 Teeth — keep a real attempts ledger. Track explicitly what you tried and exactly how it failed (the key line of the error). BEFORE each new attempt, check it against the ledger: if it's materially the same, don't run it again. This breaks the failure loop mechanically — not "I'll try not to repeat" but a ledger that forbids repetition. On a long stall, write the ledger to
.claude/idclip-log.mdso it survives across sessions.
When every wall is hard
If 2-3 real routes also fail, that's a signal: maybe the goal is wrong or a piece of info is missing. Then stop and bring it back to the user honestly — what you tried, what's left. No-clip is smart bypassing, not infinite retrying.