Build a small standalone local app to review the Amazon T-shirt research already saved in this project. Inspect the existing research format first and preserve source files. Do not connect to another app or a production service.

Start with one simple review screen: keyword/idea, Amazon link, saved page-3 total result count, checked date, and Pending / Approved / Rejected decisions. Add search, sort, undo and persisted decisions. Counts are historical snapshots, not demand or sales estimates. Test decision changes in isolated storage before touching actual user records.

After that works, add these stages one at a time and demonstrate each:
1. Approval-triggered collection of organic, non-sponsored Page 1 products. Save only verified matching products with a MAIN Clothing, Shoes & Jewelry BSR strictly below 4,000,000. Skip uncertain matches, unreadable ranks and subcategory-only evidence. Follow website constraints, use conservative pacing, save progress and exclusions, support cancellation, and stop on access challenges. Never purchase anything.
2. An app-owned background capture worker triggered by approval, with app progress and actionable errors instead of recurring no-work messages in the build chat. Make startup/shutdown and browser ownership explicit. Crop to the shirt/design image; preserve originals, BSR/category, checked date and product link.
3. Individual image approval/rejection, separate from keyword approval. Enlarge the actual artwork so a person can read its printed phrase. Listing titles are not evidence of artwork wording. Allow corrections and show review progress.
4. For this trusted personal local app, investigate the OFFICIAL local Codex CLI using the existing ChatGPT subscription sign-in for analysis. Verify current official support and usage limits. Never extract/copy/expose OAuth tokens or silently fall back to separately billed API-key usage. If the route is unavailable, explain before choosing a different provider or billing setup. Do not add credentials or paid services without direction.
5. Analyze approved artwork only: exact visible phrases, depicted characters, motifs, text-only / graphic-only / mixed classification, styles, printed ink colors excluding fabric/background, typography, composition, overall concept and uncertainties. Never reconstruct missing text from a listing title. Preserve existing analysis and allow human edits.
6. My ideas: Research → Review captures → Final report. Generate one current report per idea automatically only after collection, all image decisions, and analysis of approved images finish. Include only approved references. If everything is rejected, say no designs were selected. Update the current report when relevant decisions change instead of mixing ideas.
7. Final report with Markdown, structured JSON, a ZIP of approved reference images and Copy AI generation brief. Preserve report/image provenance. Encourage original exploration, exact spelling and one coherent direction rather than copying all references.

Keep the interface simple. Preserve original files and saved decisions during migrations. Run checks appropriate to each stage and show the real workflow before calling it complete. Do not publish, upload, buy, schedule recurring tasks or mutate real decisions for testing without explicit authorization.

After human review, the design handoff can be: paste the final brief into ChatGPT and ask, “Create five distinct T-shirt design concepts using this research brief.” Review wording and artwork, then verify export requirements. Concepts are not proof of sales or automatically cleared print-ready files.