B2.1Bloom · AnNot started

Classify failure into a 4D

Reading depth

What you'll learn

When an LLM interaction fails, the 4Ds tell you which kind of failure it is — Delegation (wrong task), Description (weak prompt), Discernment (skipped verification), or Diligence (deployment miss).

  • Delegation = wrong task for the model. Fix: change task shape (tools, decomposition, different system).
  • Description = right task, weak prompt. Fix: re-describe (role, constraints, format, examples).
  • Discernment = output trusted without verification. Fix: check facts, format, edge cases.

The 4D framework — Delegation, Description, Discernment, Diligence — is Anthropic's named taxonomy for diagnosing why an LLM interaction failed. It's not a checklist; it's a *triage* framework. When something goes wrong, naming the D tells you which lever to pull. Pulling the wrong lever (e.g. tweaking the prompt when the task is structurally undelegatable) is the canonical waste-of-day pattern.

Delegation failures: the task was a bad fit for the model in the first place. Precise multi-step arithmetic without a calculator tool, real-time data without a search tool, exact factual recall outside training cutoff — no prompt rewrites these into success. The fix is a *different task shape*: add a tool, change the surface, decompose, or hand the work to a non-LLM system.

Description failures: the task was delegatable but the prompt didn't carry enough information — vague role, missing constraints, ambiguous output shape, no examples. The output reflects the gap. The fix is to re-describe — name the role, name the constraint, name the format, add a worked example. Discernment failures: the output was checked too lazily and a wrong-but-plausible answer slipped through. The fix is verification: spot-check facts, validate format, run edge cases. Diligence failures: a deployment-side concern was missed — appropriate human oversight, bias awareness, misuse-risk planning.

The classification is structural, not by symptom. 'Wrong arithmetic' could be Delegation (no calculator) or Description (didn't ask the model to think step-by-step). 'Hallucinated citation' could be Delegation (no retrieval tool, asked to be authoritative) or Discernment (output was used without verification). Naming which D applies determines whether the fix is architectural, prompt-side, or process-side.

Key points

  • Delegation = wrong task for the model. Fix: change task shape (tools, decomposition, different system).
  • Description = right task, weak prompt. Fix: re-describe (role, constraints, format, examples).
  • Discernment = output trusted without verification. Fix: check facts, format, edge cases.
  • Diligence = deployment-side miss (oversight, bias, misuse). Fix: process and policy.
  • Triage by D before reaching for a fix. Wrong-D fix = wasted effort.

Examples

'The model fabricated a court ruling citation.'

If you asked it to be authoritative on case law without a retrieval tool: Delegation (give it search). If you asked with retrieval but skipped verification before publishing: Discernment. If both were fine but you deployed it without legal review: Diligence.

'Output is technically correct but unusable formatting.'

Description. The prompt didn't specify the output shape. Fix: add a format spec, ideally with one worked example.

Pitfalls

  • Reaching for prompt tweaks (Description fix) when the failure is Delegation. No prompt makes 'multiply 17-digit numbers exactly' work without a tool.
  • Calling everything 'Description' because it's the most familiar lever. Misclassification is the trap.
  • Skipping Diligence because the output 'looks right' — diligence failures are usually invisible until they're a news story.

Source notes: 00-academy-basics/notes/02-ai-fluency.md

Ask Claude

Build a prompt with this lesson + your question, copy it, and open the Claude Project in a new tab.