preview-pick

name	preview-pick
description	Generate cheap options first, let the user pick, then commit budget to the chosen one. Use before expensive image/video generation, when the user wants choices ("give me a few thumbnails"), or to approve a video visually before paying for clips. Keywords - options, variations, preview, storyboard, pick, choose, draft.
allowed-tools	Bash Read Write Edit Glob Grep

Preview → Pick → Commit

Don't fire-and-pay. Iterate cheaply, commit expensively. Generate low-cost options or a storyboard, show them, let the user choose, then spend on the full-quality render.

Call each workflow through the CLI: node workflows/cli.cjs <command> '<json-args>' (node workflows/cli.cjs list shows all commands).

Cost reality (per image): 512px ≈ $0.045 · 1K ≈ $0.067 · 2K ≈ $0.101 · 4K ≈ $0.15. Cheaper still: "imageModel":"lite" (Nano Banana 2 Lite) ≈ $0.0336 flat at any size — the default choice for preview/option rounds; re-render the winner with flash or pro. A video clip ≈ $0.40/s. So previewing first is real money saved, not just nicer UX.

---

Pattern A: Image options (thumbnails, hero, graphics)

# 1. Cheap preview batch (512px by default)
node workflows/cli.cjs generateImageOptions '{"prompt":"Bold thumbnail: shocked student looking at phone, neon accents, no text","count":4,"outputDir":"projects/{name}/output-contents/options","aspectRatio":"9:16"}'
# → result.data.imagePaths → SHOW all to the user as a numbered contact sheet

Show the options, ask the user to pick a number. Then commit only the chosen one:

# 2. Commit the pick at full resolution (uses it as reference to preserve the look)
# chosenImagePath is the picked option from result.data.imagePaths (e.g. index 2)
node workflows/cli.cjs finalizeImage '{"chosenImagePath":"projects/{name}/output-contents/options/option-3.png","prompt":"Bold thumbnail: shocked student… (same as above)","outputPath":"projects/{name}/output-contents/thumbnail.png","imageSize":"2K"}'
# → result.data.finalPath, result.data.cost

Total: 4×$0.045 preview + 1×$0.101 final ≈ $0.28, vs $0.60 to blindly make 4 at 4K.

---

Pattern B: Storyboard before video (approve, THEN pay for clips)

The biggest saver. Generate one keyframe per scene (~$0.067 each) and get sign-off before spending ~$0.40/s on clips.

# referenceImagePath keeps the character consistent across scenes
node workflows/cli.cjs generateStoryboard '{"scenes":[{"name":"struggle","prompt":"Student overwhelmed at messy desk, late night"},{"name":"discovery","prompt":"Same student opens the app, hopeful"},{"name":"result","prompt":"Student relaxed, clean schedule on screen"}],"outputDir":"projects/{name}/output-contents/storyboard","aspectRatio":"9:16","referenceImagePath":"projects/{name}/assets/characters/char-main-front.png"}'
# → result.data.keyframes → SHOW the whole storyboard to the user

Stop and review. Ask: same character throughout? right emotion/flow? Regenerate any weak scene. Only after approval, generate the clips with the video skill, passing each approved keyframe as the referenceImagePath (generateVideoFromImage / generateSpeakingVideoFromImage).

3 keyframes (~$0.20) approved → then ~$7 of clips, with no surprises.

---

Pattern C: Pick → refine loop ("like #2, but warmer")

Picking is rarely the end — users want to tweak the winner. Don't regenerate from scratch (you'll lose what they liked); vary the pick:

# Refine the chosen option, keeping its composition (still cheap at 512-1K)
node workflows/cli.cjs generateImageVariation '{"referenceImagePath":"…/options/option-2.png","prompt":"Same composition and subject, but warmer golden-hour lighting and a softer smile","outputPath":"…/options/option-2b.png","imageSize":"512"}'

Loop: show → adjust one thing per round (lighting, expression, background, crop) → regenerate the variation. When the user says it's right, finalizeImage that file at full resolution. Two or three refine rounds still cost less than one blind 4K miss.

Keep every round's files — users often walk back ("actually 2b was better"). Number them option-2b, option-2c, … so picks stay unambiguous.

---

Pattern D: A/B variants (marketing tests, not indecision)

When the user wants to test content (not just choose), deliver true variants:

1. Vary ONE deliberate dimension per variant — hook, thumbnail emotion, CTA, opening scene — and say which. Same everything else, so the test is clean. (For copy: generateHooks already returns multiple angles; pick 2 distinct ones.)
2. Produce both finals (A and B) and package both: …/publish/variant-a/instagram/, …/publish/variant-b/instagram/.
3. Record what differs in each variant's post.json via the caption/notes and tell the user what to measure (CTR for thumbnails/hooks, watch-through for opening scenes).

Cost doubles on the final render only — previews already existed from Pattern A.

---

Pipeline-first + audit trail (mandatory)

• Author <content-id>.pipeline.json IN the content folder BEFORE any paid round (nodes = CLI commands, {{node.data.field}} refs — e.g. options → finalize wired together), then node workflows/cli.cjs runPipeline @<file>. See workflows/pipelines/README.md.
• Keep a prompts.txt in the content folder: the option-round prompt, each refine round's prompt, and the finalize prompt — every round gets a one-line RESULT note.
• Log every paid round via createGenerationManifest / addManifestEntry (AGENT-GUIDE Step 5), including the option rounds — cheap still counts.

---

When to use

• "Give me a few options / variations" → Pattern A
• Any expensive image (hero, 4K, ad creative) → Pattern A (preview at 512 first)
• Any multi-clip / character video → Pattern B (storyboard before clips), then the video skill
• "Close, but change X" after a pick → Pattern C (vary the pick, don't restart)
• "Which hook/thumbnail works better?" → Pattern D (clean A/B, package both)
• One-off cheap asset where the user trusts a single gen → skip; generate directly

Always state the cost of the cheap step and the commit step so the user chooses with eyes open.