Skip to content

Chapter 5 — Quality & Prompt Tuning

You've run workflows, made lightweight variants, and used PatchWork's creative tools. By now you've also seen generations that aren't quite right — a face that doesn't match, lighting that's off, a hand pose that looks stiff. This chapter is about what to do about it.

The short answer: most of the time, the system handles quality automatically. The longer answer is what this chapter covers.

Chapter at a glance

flowchart TD
    A[Candidate set<br/>arrives] --> B{All 4 candidates<br/>have the same issue?}
    B -->|No| C[Pick the<br/>strongest take]
    B -->|Yes| D[Prompt-tuning<br/>skill]
    D --> E[Iteration loop:<br/>generate, judge, adjust]
    E --> F[Dialed-in<br/>prompt]
    F --> A

When to trust the system, when to step in, how to step in.

What you'll learn

  • How the system handles quality automatically — what the auto-QA loop catches, what the pre-generation sanity check filters, when the Generation Runner reruns on your behalf
  • When you actually need to step in — the difference between "let it retry" and "this needs the prompt-tuning skill"
  • The prompt-tuning skill — what it is, when to invoke it, what to expect from the loop, what it costs
  • Judging candidates — picking the strongest take from the 4 you're shown, knowing when to accept "good enough" vs. push for more

Before this chapter

Read Chapter 4 — PatchWork Creative Tools. You should have generated a few workflows by now and seen what good and not-so-good candidates look like in PatchWork's gallery.

Sections

  1. How quality is handled — what the system does automatically, so you know when you actually need to intervene
  2. Prompt tuning — the skill in practice: when to use it, how to invoke it, what comes back
  3. Judging candidates — picking the strongest take from a candidate set, knowing when to ship vs. iterate

Try this

Pick a candidate set in PatchWork — ideally one from the workflow you ran in Chapter 2.

Steps:

  1. Look at the 4 candidates side by side. Pick the strongest in 5 seconds, then write down (in plain text, one sentence) why you picked it (Judging candidates)
  2. Find one candidate set in the workflow where all 4 candidates share the same issue (lighting wrong, wardrobe drifted, pose stiff). If none have an all-4 issue, look harder — most workflows have at least one
  3. For that bottleneck scene, ask Claude to invoke the prompt-tuning skill: "Tune the prompt at <path>. Reference is <reference-image>. The lighting keeps coming out too cool. Try 5 iterations." (Prompt tuning)
  4. Watch the iteration loop. Read the candidates that come out at each step. Confirm the loop converged

If you hit:

  • Candidates are bad but you can't articulate why → re-read Judging candidates, specifically the priority order (face → composition → wardrobe → pose → setting)
  • Auto-QA flagged a candidate that looked fine to you → it probably caught something at zoom level you'd miss. Trust the system. See How quality is handled
  • Prompt-tuning loop ran but the result is still wrong → the issue may not be prompt-fixable. Re-read the "When prompt-tuning isn't enough" section of Prompt tuning
  • You're tempted to tune every scene, not just the bottleneck one → don't. See targeted investment in Judging candidates
  • Garbled background text or signage is bothering you → ignore it. See "What NOT to look at" in Judging candidates

When you're ready

→ Next: Chapter 6 — Lvl 3-4 Variants. Now that you can judge a generation and reach for the prompt-tuning skill when you need to, you're ready for bigger variants — environment changes, camera changes, structural changes. These go through a testing phase, and the testing phase leans on what you learned here.