B2.4Bloom · ANot started

Run description ↔ discernment loop

Reading depth

What you'll learn

When the output disappoints, ask 'what did the prompt leave the model to guess?' and add it to the prompt. That's the loop.

  • Description and Discernment are a loop, not a sequence.
  • Disappointing output → ask 'what did I leave to guess?' → re-describe.
  • Retrying without re-describing is sampling, not iteration.

Description and Discernment work as a loop, not as separate steps. You write a prompt (Description), evaluate the output (Discernment), find a gap, and the gap *almost always* points back at a missing piece of the prompt — not at a need to retry, raise temperature, or escalate to a bigger model. One iteration of this loop is more valuable than three retries with the same prompt.

The mechanic: when output disappoints, ask 'what did I leave the model to guess about?' Format inconsistency → format wasn't specified. Wrong tone → role wasn't specified. Missed an edge case → constraint for the edge case wasn't named. Hallucinated a fact → either no grounding source was provided (Description) or the model wasn't told to flag uncertainty (Description). The prompt is the lever; running the loop disciplines you to use it.

The wrong reflex when output disappoints: retry with the same prompt, hoping for a luckier sample. Retrying without changing the description is sampling variance, not learning. Even when retry produces a better answer, you've encoded nothing — the next run regresses. The loop produces durable improvements.

What ends the loop: when the gap you observe in Discernment can no longer be closed by a Description change. That's the signal you've crossed into Delegation territory (the task itself was wrong) or Diligence territory (the deployment process needs a checkpoint). The loop is also a triage mechanism — it tells you when prompt iteration has stopped paying.

Key points

  • Description and Discernment are a loop, not a sequence.
  • Disappointing output → ask 'what did I leave to guess?' → re-describe.
  • Retrying without re-describing is sampling, not iteration.
  • When re-description stops helping, you've hit Delegation or Diligence.

Examples

Output: occasionally invents action items the transcript didn't contain.

Discernment found the gap. The Description re-write: add an explicit constraint, 'Do not invent action items; if no commitment was made, omit the section.' Re-run; verify; either the gap closes (loop succeeded) or it doesn't (Delegation question — maybe the input is too noisy).

Output: tone is too casual for an executive summary.

Description gap: role unspecified. Re-write: add 'You are writing for a CFO; precise, concise, no hedging.' Loop ends quickly.

Pitfalls

  • Retrying with the same prompt and counting the better sample as 'fixed.' Sampling variance ≠ improvement.
  • Skipping the gap-naming step. 'Output isn't great, let me rewrite the whole prompt' is rarely as useful as 'output is missing X, let me add the X-spec.'
  • Not knowing when to stop. After 3 loops with no movement, name as Delegation or Diligence and change tactic.

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.