Improving Your Prompt & Display Template with Feedback

Use real results and feedback traces to make your analysis smarter (build V3.5.1.0)

Prerequisite.

A working SKU Manager setup with AI analysis running. If you’re not there yet, start with the Quick Start Guide.

Your first prompt won’t be perfect — and that’s expected. The real power of SKU Manager is the feedback loop: run searches, spot issues, capture them as traces, fix the prompt, repeat. Even 80% accuracy on day one is valuable; you get to 90%+ by iterating.

Table of Contents #

1. How the Improvement Cycle Works
2. Know What to Look For
3. Submit Feedback in SKU Manager
4. What’s in a Trace
5. Bring the Trace to Your AI Project
6. Common Improvement Requests
7. Load and Test the Update
8. The Improvement Mindset

How the Improvement Cycle Works #

Run searches → spot issues → capture feedback (trace) → fix in your AI project → update SKU Manager → test again.

Each pass through the loop makes the prompt smarter and the template more useful. Think of it like training a new employee — you show them examples of right and wrong, and they get better.

Know What to Look For #

Prompt Issues — the AI is Thinking Wrong #

• Wrong product identification — says Product A when it’s Product B
• Wrong counts — “3 items” when the listing shows 5
• Math errors — price calculations that don’t add up
• Wrong decisions — a clear pass called a buy, or the reverse
• Missing information the listing clearly contained
• Bad condition calls — missed damage, misread sealed status

Template Issues — the AI is Thinking Right, but Showing it Wrong #

• Data not appearing where expected
• Colors or badges showing the wrong status
• Broken or hard-to-read layout
• Fields showing undefined or null instead of values

Patterns Beat One-offs.

Keep a notepad open while you search. “Keeps missing quantity on lot listings” is far more fixable than a single odd result.

Submit Feedback in SKU Manager #

When you find an issue, don’t just move on — capture it. Feedback creates a trace: a complete snapshot of that analysis, and the fastest route to a fix.

1. Select the listing (with its analysis showing) and click Feedback — top right of the AI Sku Analysys panel.
2. Rate the result — a row of labelled faces.
3. Pick the issue type: Wrong Calculation · Wrong Product · Missing Info · Display Error · Prompt Issue · Other.
4. Write a note: what the AI got wrong, what the correct answer is, and where the correct info was (title? description? images?).
5. Submit — the trace is written with your rating saved in it.

Specific Notes Get Fast Fixes.

❌ “Wrong quantity”  →  ✅ “AI said 2 items but the title says ‘Lot of 5’ and the photo shows 5 boxes clearly.”

What’s in a Trace #

Open traces with Data → AI Sku Analysis → Open AI Traces — each submission gets its own dated folder, newest last. Inside:

The minimum for a fix is AI_Feedback.txt + AI_RequestPayload.json — the prompt and the data are both in the payload. There is no separate prompt or template file: both live inside <Profile>_AI.json.

If the Issue Involves Images, Include the Image Files.

Misidentified product, missed damage, unread label — without Image_1.jpg etc., your assistant can’t see what the AI saw and can’t diagnose the visual call.

Bring the Trace to Your AI Project #

The fix happens in the same SKU Manager Assistant project you built the prompt in — that’s what keeps fixes consistent with everything that came before.

Option A — Quick Fix (Single, Clear Issue) #

The analysis is working but I found an issue.

Issue: [what went wrong]
Correct answer: [what it should have been]
Where the data was: [title / description / images]

Can you update the prompt to handle this correctly?

Option B — Trace-Based Fix (Complex or Unclear Issue) #

I have a trace from SKU Manager showing an analysis error.
The AI got [X] wrong — it should have been [Y].

Files attached:
- AI_Feedback.txt (my notes on what's wrong)
- AI_RequestPayload.json (the prompt and the data that was sent)
- Image files (what the AI saw)

Can you find the root cause and propose a targeted fix —
before/after, not a full rewrite?

The assistant compares what the AI said against what it should have said, checks whether the right data was even sent, finds the prompt section responsible, and shows you a before/after fix with the reasoning.

Option C — Batch Several Issues #

I've been testing my prompt and found several patterns to fix:

1. [Issue]: AI keeps [doing X] when it should [do Y]. Example: [scenario]
2. [Issue]: [description]. Example: [scenario]

Can you review my prompt and address all of these?
Show me the changes for each one.

Common Improvement Requests #

Prompt-side #

• New product variations: “My prompt handles [A] well but I also buy [B] — it’s treating [B] as invalid. Add it to the validation and pricing logic. My buy prices: […]”
• Count detection: “When the title says ‘Lot of X’ but the description mentions individual items, the AI uses the wrong number. The rule should be: […]”
• Condition calls: “A listing had [detail] visible in the photos but got rated [wrong]. It should check for [what] and rate it [correct].”
• Decision thresholds: “It recommends buy above [X]% margin; I need at least [Y]%.”
• Edge cases: “When [edge case] happens, the AI should [behavior]. Example: [listing].”

Template-side #

Presentation fixes are usually quicker. Paste your current template plus an example of the JSON the prompt returns, and ask for: a section for a new field, corrected color logic, a reorganized layout, or graceful fallbacks so missing values never show as null.

Load and Test the Update #

1. Open Data → AI Sku Analysis → AI Settings and select your profile.
2. Replace the System Prompt (AI Configuration tab) and/or the Display Template — select all, delete, paste the whole updated version from your master file.
3. If a fix added or renamed AI columns, tick them on the Fields & Columns tab.
4. Save with Start Service [Save Configuration].
5. Re-test the listing that had the problem, then 10–20 others to catch regressions.

If the fix introduced a new problem: “The fix for [original issue] works, but now [new issue] is happening. Can you adjust without breaking the original fix?”

The Improvement Mindset #

• Prioritize by impact. High-volume issues, costly wrong decisions, and easy wins first. Batch rare edge cases for later.
• Version your work. Keep the master prompt and template as text files — v1, v2, v3 — and keep the previous version as a rollback.
• Track changes. A one-line log per version (“v3: fixed lot counting, added damage flag”) pays off months later.
• Know when to stop. 90%+ on the listings you see daily is a great target; rare edge cases are fine to review by hand. The jump from 90% to 95% costs more than 50% to 90% did.

Tips #

• Use your microphone. Talking through what’s wrong is often faster than typing — Windows: Win + Shift + H starts voice typing in any text field.
• Keep traces for important fixes — they’re documentation of what was wrong and how it was fixed.
• Don’t rewrite the whole prompt for one issue. Targeted fixes are safer; your assistant should show exactly what changed and why.
• Test with real listings — real eBay data has edge cases you won’t invent.

See also: Quick Start Guide · Full Setup Guide · Setting Up AI Columns · Building Your AI Analysis