Building Your AI Analysis

Component	What it does	Where it lives in SKU Manager
System prompt	Instructions telling the AI how to analyze listings	Your AI profile → AI Configuration
Display template	HTML that formats the result in the AI Sku Analysys panel	Your AI profile → Display Template

The prompt outputs JSON data → the template displays it visually. The two are built together so the field names match.

Once set up, your AI project becomes your ongoing workspace — you come back to it to add features, tighten rules, and fix mistakes using feedback traces.

---

Before You Start #

• SKU Manager configured — whitelisted, AI Settings opened once so Python has installed. See the Quick Start Guide.
• An AI assistant account — Claude (claude.ai, Pro recommended) or ChatGPT.
• Know your products — what you’re searching for, what qualifies, what disqualifies.

---

Set Up Your AI Project #

About 10 minutes, one-time.

Already Have the Project?

This is the same SKU Manager Assistant project used across all SKU Manager articles, including Internal Endpoint. If you’ve built it once, it covers everything — skip to Describe What You’re Looking For.

Using Claude #

1. Download the SKU Manager Knowledge Package and unzip it.
2. Go to Claude.ai → Projects → Create Project. Name it SKU Manager Assistant.
3. Open the unzipped folder, select all the files (Ctrl+A), and drag the whole selection onto the project’s Context panel in one go — the project won’t accept a folder.
4. Start a chat inside that project, using the latest model with extended thinking.

Using ChatGPT #

Open our SKU Manager Prompt/Template Assistant Custom GPT — the same files are already loaded.

What’s in the knowledge files? #

They teach the assistant how prompts and display templates connect, which eBay fields exist, how AI columns work, worked examples from different kinds of buying, and how the feedback/trace system feeds fixes back in. That means it can create a prompt and template from scratch, keep them in sync, and — later — read your traces and make targeted fixes.

---

Describe What You’re Looking For #

About 30–60 minutes for a first working version.

Start a new chat in the project. The goal is to explain your buying process the way you’d explain it to a person. Start simple:

I want to create a new prompt and display template.

I search for [SEARCH TERM] and see [TYPES OF PRODUCTS] in the results.
When I'm going through listings, here's what I'm looking for:
- [What makes something a yes]
- [What makes something a no]
- [Things I can only tell from pictures]
- [Things I need to read in the description]

Let's start simple, just to get it working.

Describe your mental process #

• Product identification — “I’m looking for [specific brand/model], not [other things that show up in the same search].”
• Condition requirements — “It needs to be [sealed / working / no damage].”
• Visual inspection — “From pictures, I’m checking for [cracks / missing parts / box damage].”
• Description details — “In the description, I look for [keywords / red flags].”
• Price tiers — “I’ll pay $X for [condition A] but only $Y for [condition B].”

Example conversation #

I want to create a prompt and display template.

I search for "digital camera" and I'm trying to buy working, undamaged
cameras at a good price.

When I look at listings, I check:
- Is it the kind of camera I buy (a real brand, not a toy or an accessory)?
- Does it power on? Does the seller say it's tested?
- Any damage I can see in the photos — cracked screens, dents, lens scratches?
- Does it accept returns? I avoid no-return listings on untested items.

From the pictures I'm looking at body condition and the lens glass.
Some sellers don't show everything clearly.

Let's start simple — just help me identify the right products and flag
any obvious damage.

What the assistant will do #

Ask clarifying questions if it needs to, then generate a prompt with a proper JSON structure and a matching display template, with the field names aligned between the two. Copy both into a text file where you can get at them again.

Mindset.

Your first goal isn’t perfection — it’s the full round-trip working. Refine after you see real results.

---

Load It Into SKU Manager #

About 5 minutes.

Both pieces go into one AI profile: open Data → AI Sku Analysis → AI Settings, paste the prompt into System Prompt on the AI Configuration tab and the template into the Display Template tab, then save. On a first run this is the 3-step wizard; the full walkthrough — including the trigger filter, which is a separate step the wizard does not cover — is in the Quick Start Guide.

Paste Whole, Every Time.

The code boxes can truncate what you paste. Keep the master copy in your text file, never hand-edit inside the box, and re-paste the whole prompt or template on every change. Start Service [Save Configuration] is the save button when you’re editing an existing profile.

---

Improving Your Prompt Over Time #

When the AI gets something wrong #

1. Select the listing (with its analysis showing) and click Feedback — top right of the AI panel.
2. Rate the result, pick an issue type, and describe the problem in plain language: “This should have been flagged as damaged”, “It missed that this is a lot of 3”, “The model identification is wrong — it’s an X, not a Y.”
3. Submit. That writes a trace — a complete snapshot: the data sent, the exact prompt, the AI’s response, the images, and your notes.
4. Open the trace via Data → AI Sku Analysis → Open AI Traces, and bring the files to your AI project — at minimum AI_Feedback.txt, AI_RequestPayload.json, and any images:

I have feedback on my prompt. The AI got something wrong.
[Attach the trace files]
Here's what happened: [your notes]

Please identify which part of the prompt caused this and propose a
targeted fix — before/after, not a full rewrite.

5. Apply the updated prompt to your profile and re-test the same listing to confirm the fix.

Adding features later #

My prompt is working. Now I want to add:
- [New feature, e.g. price calculations]
- [New feature, e.g. AI columns for the grid]

The assistant knows whether a change needs the prompt, the template, or both.

---

Adding AI Columns #

AI columns put key decisions straight into the results grid — short, exact values like Can Return: Yes — so you can scan, sort and filter without opening each listing. Many users show the same key values in both places: columns for scanning, the template for the full picture.

Columns are Defined in the Prompt, Not in the App.

You don’t create columns in SKU Manager first. Ask your assistant to add them to the prompt’s ## AI COLUMN CONFIGURATION section — the app then auto-detects them from the prompt.

1. In your AI project: “Add an AI column called Can Return that returns exactly one of: Yes, No, Unknown.”
2. Paste the updated prompt into your profile (whole prompt, as always).
3. On the Fields & Columns tab, your new column appears under Detected — not selected — tick it, and ignore the junk keys the detector also lists.
4. Save. The column fills in on the next run.

Exact Values Matter.

The grid sorts and filters on these values verbatim. Keep spelling and capitalization consistent (Yes, not yes or Y), give every column an Unknown fallback so it’s never blank, and make sure values can’t overlap.

---

Include Reasoning in Your Prompts #

Ask your assistant to include reasoning and source attribution in the output — a natural-language explanation of how the AI reached each conclusion, and which fields and photos informed it.

Include a detailed reasoning section that explains in natural language
how you arrived at each major conclusion. Also include source attribution
showing which fields and photos informed each decision.

What that looks like in the output:

• Reasoning: “Condition is graded Good based on the photos: the body shows light wear but the lens glass is clean in photos 3–4. Can Return is Yes — the seller lists a 30-day return policy.”
• Attribution: “Model: Title, Description. Defects: Photos 2–4. Can Return: Returns field.”

This tells you why the AI decided what it did and which data caused a wrong call — which makes feedback traces far faster to act on.

---

Tips for Success #

• Start simple — get basic identification working before adding price logic or columns.
• Describe your thinking — the more of your mental process you explain, the better the prompt.
• Pictures matter — tell the assistant what you look for in photos.
• Test with real listings before tweaking.
• Use feedback traces — don’t guess at what went wrong; capture the actual data.
• Keep everything in one project — the assistant builds on context from earlier chats.
• Version your work — keep the master prompt and template in text files, saved as v1, v2… before big changes.

Questions? Live Support at uBuyFirst.com, or su*****@*******st.com.

See also: Quick Start Guide · Full Setup Guide · Using Feedback & Traces