Don't be Dumb!

Coding agents often fall into reasoning tunnels, digging deeper into speculative fixes while ignoring the root cause. This leads to broken builds, regressed code, and wasted tokens. This skill provides a structured recovery layer to halt speculative edits and force a reality check based on evidence rather than assumptions.

What it does

• Stops speculative loops by halting multi-variable edits and workarounds when progress stalls.
• Establishes Last-Known-Good states by auditing git history, logs, and user confirmations to find the boundary of failure.
• Audits evidence by strictly separating established facts from unproven assumptions and citing sources.
• Challenges hypotheses via internal self-questioning to detect confirmation bias or symptom-fixing.
• Plans minimal diagnostics by identifying the smallest action needed to prove or falsify the current direction.

How it works

1. Trigger the skill manually with don't be dumb or allow the agent to auto-activate when it detects repeated failed attempts.
2. Reconstruct the original task and constraints to strip away accidental scope creep.
3. Analyze the gap between the last working state and the first failure to isolate the actual change that caused the regression.
4. Execute a diagnostic plan that prioritizes gathering evidence over modifying code.

Frameworks & tools

This skill is platform-agnostic and works with any environment where the agent has access to git, test runners, and terminal output. It is designed for use in Claude Code, Cursor, and other agentic IDE environments.

Why this beats prompting it yourself

Manual prompting often fails once the agent is already in a reasoning loop because the context is saturated with its own incorrect assumptions. This skill provides a rigid, phase-based framework that forces the agent to move out of its own context and treat its previous actions as untrusted data.

Use cases

• Recovering when an agent keeps "fixing" the same file but the test still fails.
• Identifying why a previously working feature suddenly broke after a series of complex refactors.
• Breaking out of a loop where the agent keeps suggesting the same library update or config change.

Known limitations

Does not perform automated rollbacks without explicit permission. Requires access to local history or git to establish a credible last-known-good state.