📁 E-commerce
Product Return Reason Cluster from Review Export (Defect vs Sizing vs Expectation)
Cluster return reasons into defect, sizing and expectation so you fix the product, not the copy
0Reviews
Prompt
Act as a returns analyst who reads only the pasted review and returns export. You cluster return reasons into defect, fit and sizing, and expectation, then say what evidence supports each cluster. You never invent a return rate, a defect rate, a cost figure, or a customer quote that is not in [Export]. You do not blame the customer for a product problem. You work only from Inputs. Do not invent stats, citations, quotes, URLs, names, IDs, or records that are not in Inputs. Inputs: - Review or return export with reason text and dates: [Export] - Product name and variant list (size, color, model): [Variants] - Product facts I can support: materials, measurements, care, intended use: [Facts] - Listing claim set (what the page promises): [Claims] - Period covered: [Period] - My cost per return, if I share it: [CostPerReturn] - Output format: [Format] - Language: [Lang] Generate: 1. Export ledger: how many return lines and review lines, which reasons appear, and how many are blank. Print the counts. Forbidden: invented totals, invented rates. 2. Cluster pass: group reasons into four buckets - defect, fit and sizing, expectation, and other. Quote one representative line per bucket and give the line count. Label every cluster with its n. 3. Defect analysis: for defect clusters, list the failure mode and which part, mapped to [Variants] if the export names one. If the export does not name the part, write NOT IN INPUTS. 4. Fit and sizing analysis: what dimension or size is named, how the listing states it from [Claims], and where the stated measurement lives in [Facts]. If no measurement exists in [Facts], write MISSING MEASUREMENT rather than guessing a size run. 5. Expectation gap: the claim from [Claims] the customer expected, the line that shows the gap, and whether [Facts] supports the claim. Unsupported claims are flagged, not fixed. 6. Fix priority: ranked by (cluster line count) times (whether [CostPerReturn] or [Facts] can price it). Show the equation. Any dollar figure uses [CostPerReturn] only. 7. Copy changes: the listing lines that must change, each with the exact replacement wording that stays inside [Facts]. 8. Gaps: what the export does not show and what I still owe you. Constraints: - Every cluster cites a quoted line from [Export]. Quote or drop. - No invented rates, no invented defect counts, no invented savings, no blame on the buyer. - Honor [Period], [Format], and [Lang]. - Missing-data policy: if a field was blank, write NOT IN INPUTS. - End with a five-bullet list of what the user still owes you.
Instructions
Paste the return and review export exactly as exported, one entry per line. Replace every [bracket] before running. Works on ChatGPT, Claude, and Gemini. Blank fields become NOT IN INPUTS, so the model flags a gap instead of inventing a rate.
Examples
Example Input
Export (returns, 09/01 to 09/30): 09/03 "zipper broke after two wears, black, M" (return, defect) 09/05 "runs small, ordered M, usually wear M" (return, sizing) 09/08 "fabric thinner than it looks on the page" (return, expectation) 09/09 "stitching came loose at the pocket, black, L" (return, defect) 09/12 "color darker than the photo" (return, expectation) 09/15 "sleeves too short, navy, L" (return, sizing) 09/18 "beads fell off in the wash, ivory, S" (return, defect) 09/22 "expected heavyweight, this is light" (return, expectation) 09/26 "length is shorter than the chart says" (return, sizing) 09/29 "button missing on arrival, black, M" (return, defect) Reviews (09/01 to 09/30): 09/04 5 stars "fits perfectly, third one I bought" (M, black) 09/11 4 stars "lovely but the zip sticks" (L, navy) 09/19 5 stars "true to size, great autumn weight" (S, ivory) 09/24 3 stars "photo makes it look heavier than it is" (M, black) Variants: S, M, L; colors black, navy, ivory Facts: Cotton shell, viscose lining. Care: cold wash, do not tumble dry. Intended use: autumn jacket. Beaded trim is glued. No published length or chest measurement on the page. Claims: "Heavyweight autumn jacket", "Rich true-to-photo colour", "Runs true to size", "Beaded trim" Period: 09/01 to 09/30 CostPerReturn: 11 Format: ledger + cluster table + copy changes Lang: English
Example Output
1. Export ledger.
Return lines: 10. Review lines: 4. Blank reason lines: 0. Reasons present: defect (4), sizing (3), expectation (3).
2. Cluster pass.
| Bucket | n | Representative quoted line |
| Defect | 4 | "zipper broke after two wears, black, M" (09/03) |
| Fit and sizing | 3 | "runs small, ordered M, usually wear M" (09/05) |
| Expectation | 3 | "fabric thinner than it looks on the page" (09/08) |
| Other | 0 | none |
3. Defect analysis.
- Zipper failure, 09/03, black, M. Part: zipper. [Variants] names black and M.
- Stitching loose at pocket, 09/09, black, L. Part: pocket stitching.
- Beads fell off in the wash, 09/18, ivory, S. Part: beaded trim. [Facts] says beaded trim is glued, so washing failure is consistent with the construction, not with misuse.
- Button missing on arrival, 09/29, black, M. Part: button, arrival defect. Which step failed: NOT IN INPUTS.
4. Fit and sizing analysis.
Named dimensions: sleeves too short (09/15, navy, L), length shorter than the chart says (09/26). [Claims] states "Runs true to size". [Facts] publishes no length or chest measurement on the page, so the chart the customer compared against is MISSING MEASUREMENT. No size-run recommendation can be derived. Three of four positive signals in the reviews ("fits perfectly", "true to size") sit in M and S, and both sizing complaints are L, but n is too small to change the chart, and no measurement change is made.
5. Expectation gap.
Claim "Heavyweight autumn jacket": lines "fabric thinner than it looks on the page" (09/08) and "expected heavyweight, this is light" (09/22), plus review "photo makes it look heavier than it is" (09/24). [Facts] gives cotton shell with viscose lining and an autumn intended use, so the claim is unsupported at the level the customer expected. Flagged.
Claim "Rich true-to-photo colour": line "color darker than the photo" (09/12). [Facts] names no colour reference, unsupported. Flagged.
6. Fix priority.
Equation: priority = cluster line count x whether the fix is priceable from [CostPerReturn]. 11 per return applies to every return, so it does not separate clusters.
1. Defect, 4 returns x 11 = 44 in [Period]. Fix is construction: zip spec, pocket stitching, and glued-bead trim; the bead spec is already visible in [Facts].
2. Expectation, 3 x 11 = 33. Fix is copy, priceable without [CostPerReturn].
3. Fit and sizing, 3 x 11 = 33. Fix needs a real measurement, which is MISSING MEASUREMENT, so it is blocked, not priceable.
Total [Period] return cost from this export: 10 x 11 = 110.
7. Copy changes.
- "Heavyweight autumn jacket" becomes "Cotton-shell autumn jacket with viscose lining" (inside [Facts]).
- "Rich true-to-photo colour" becomes "Colour shown in studio light; see the fabric swatch photo" only if the swatch photo exists, otherwise cut the line. [Facts] names no colour reference, so the claim is cut, not softened.
- "Runs true to size" is held until a measurement chart is published; it cannot be supported or refuted from [Facts].
8. Gaps.
The export does not name which batch or factory date the defect pieces came from, and it does not include the measurement chart.
What you still owe me:
- The measurement chart you intend to publish, with length and chest per size
- Batch or PO numbers for the four defect returns
- Whether the swatch photo exists for the colour claim
- Return rate denominator: units shipped in [Period], which is NOT IN INPUTS
- Cost per return for exchanges, if it differs from 11