Practical Lesson 6 of 7

When Claude Gets Stuck

Stall patterns occur when Claude Code enters a cycle of repeated errors or produces implementations that diverge from requirements despite multiple correction attempts. The instinctive response — adding more explanation — is generally counterproductive. This lesson documents diagnostic indicators and effective recovery strategies.

Heuristic: if three successive explanations fail to resolve an issue, the problem lies in the approach, not the explanation. A change in strategy is required.

Diagnostic Indicators

The following patterns signal a stall condition:

  • Identical errors recurring after ostensible corrections
  • Disproportionate time spent explaining rather than implementing
  • Confidently generated but architecturally incorrect output
  • Repetitive acknowledgment of errors without behavioral change

Recovery Strategies

Context Reset

Execute /clear to discard accumulated context. Stall patterns frequently originate from context degradation rather than task complexity. Re-introduce only the essential specification in the fresh session.

Task Decomposition

Reduce scope to an isolable subproblem. If "build the auth system" fails, "validate the email format" establishes a working foundation from which complexity can be incrementally added.

Example-Driven Specification

Provide a minimal working implementation as a reference. Concrete code communicates intent with higher fidelity than verbal description and anchors the model to a known-correct pattern.

Lexical Reframing

Alternative terminology activates different regions of the model's learned patterns. "State machine" versus "handle transitions" may produce fundamentally different implementation approaches.

Model Escalation

If a task stalls on Sonnet, switching to Opus provides access to deeper reasoning capabilities. Conversely, overly complex reasoning on a simple task may benefit from Haiku's more direct generation patterns.

Early Detection

The critical meta-skill is recognizing stall patterns before significant time is invested. The three-explanation threshold serves as a practical decision boundary: if three attempts at clarification fail to produce correct output, the current approach should be abandoned in favor of one of the recovery strategies above.

Additional explanation within a degraded context is unlikely to resolve the issue. Effective recovery requires a structural change: different framing, reduced scope, fresh context, or a reference implementation that bypasses the communication failure.

Key Takeaways

  • Three failed explanations indicate a strategy problem, not a communication problem
  • /clear resolves context-degradation-induced stalls
  • Task decomposition isolates the failure to a tractable subproblem
  • Concrete examples outperform verbal descriptions for specification
  • Model escalation (Sonnet to Opus) provides access to deeper reasoning