Paste This Codex Prompt. Let It Do Your T-Shirt Research.

Act as a print-on-demand keyword research analyst. Execute this research and deliver a genuine downloadable Excel workbook. Do the work; do not stop at a plan or instructions.

This is a standalone task. Use Codex's built-in, agent-controlled browser and available workspace file-generation tools, following their operating instructions. Do not depend on POD Command Center, a database, private project code, a particular plugin, operating system, filesystem path, paid subscription, or API key. Use the defaults below without asking routine setup questions.

MANDATORY BROWSER CHOICE — CODEX'S BUILT-IN BROWSER, IN THE BACKGROUND
- Use Codex's built-in/in-app agent-controlled browser for research. This browser is explicitly allowed even when the Codex app runs on my computer. A separate cloud machine is NOT required, and you must not stop merely because remote hosting cannot be verified. Describe the browser accurately; do not claim it is a cloud computer without evidence.
- Keep research tabs in the background. Do not show or focus a browser pane, open a visible preview, or launch an external browser window. Do not open, inspect, navigate or control my personal Chrome, Edge, Safari or other personal browser tabs. Do not use browser-extension access to those browsers or native desktop/mouse/keyboard control.
- Select the built-in/in-app browser explicitly using the available tool's documented instructions. When the documented cua API is available, use the "iab" browser with visible: false when creating a research tab. Use the supported equivalent in another tool environment; these identifiers are guidance for tools that actually expose them, not an assumption that every session has that API. Never choose an unspecified/default browser that could route to my personal browser.
- A dedicated hosted agent browser is also acceptable if already available and usable without opening or controlling my personal browser. Do not impose a remote-only prerequisite or change this task into a cloud-environment setup task.
- If the available tools cannot provide either permitted browser in the background, stop and state that specific missing capability. Do not substitute my personal browser, install an extension, or launch a separate browser through shell commands as a workaround. Search snippets alone cannot replace interactive Page 3 verification.
- Creating the workbook and logs with the session's normal workspace/file tools is allowed, including a local task workspace. Return file links without automatically opening the workbook, browser previews or other output files. I will choose whether to open them.

DEFAULTS
- Marketplace: Amazon.com, United States, English.
- Product: graphic or typography print-on-demand T-shirts.
- Seed: one word supplied by the user, or choose one ordinary action, hobby, occupation, recipient, or occasion word yourself. Avoid protected names. Search KeywordTool for "<seed> shirt".
- Work on ONE seed at a time. Do not begin with a multi-seed discovery campaign.
- Check up to 40 unique eligible keywords for the current seed.
- Only if that seed produces ZERO qualifying keywords in those checks, try a different seed.
- Maximum: 3 seed attempts and 120 unique candidate validations TOTAL per run. These are limits, not minimum targets. Do not reset the total counter when changing seeds.
- Competition requirement: fewer than 1,000 results according to the verified PAGE 3 results-total heading.
- Output: a new timestamped folder in the current writable task workspace. Do not overwrite earlier research.

1. CHECK CAPABILITIES
Select a permitted background browser as described above, then confirm that interactive public browsing and workspace file creation are available. Do not require proof of remote hosting. Record the browser/tool used, start time, settings and any visible Amazon delivery-location setting in a run log in the task workspace. Use UTC for observations.
If a required capability is missing, save whatever evidence exists, explain the specific limitation, and finish honestly as blocked or partial. Search-engine snippets cannot prove an Amazon Page 3 count. Never pretend to have browsed or created a file.

2. COLLECT REAL SUGGESTIONS FOR ONE SEED
Use KeywordTool.io's Amazon mode, United States, English. Preserve the exact seed query, source URL, observation time and actual suggestion wording.
Collect the suggestions exposed through the normal public interface. Use ordinary Load More controls when needed. Stop collecting once at least 40 unique eligible candidates are available, or when the accessible suggestion list is exhausted. A loaded batch may contain more than 40 suggestions; preserve all observed suggestions, but validate no more than 40 for that seed.
Do not invent suggestions, pad the list, infer phrases from search snippets, or treat locked metrics/placeholders as data. No search-volume metric is required.
If KeywordTool requires unavailable access, stop and disclose that limitation. Do not silently substitute another source. Lack of access is not evidence that a seed produced zero results.

3. FILTER AND DEDUPLICATE
Preserve each exact original phrase. Deduplicate using Unicode NFKC normalization, lowercase, trimming and collapsed whitespace; do not rewrite the phrase used for its Amazon search.
Keep phrases with relevant T-shirt/design intent. Exclude irrelevant products, sizing-only noise, malformed listing fragments and unrelated garments or technical features.
Exclude clear brand, franchise, character, celebrity, named-person, school, team, band, film/TV, organization, song-title and lyric references. Hold uncertain associations outside the final workbook; record a specific reason and consult sources when useful. A quick screen is not trademark clearance.
If prior research is supplied, use its canonical keys to avoid duplicate findings and skip phrases validated within 90 days unless a refresh is requested. Otherwise deduplicate within this run and state that historical deduplication was unavailable.
Prioritize specific relevant phrases without unresolved risk flags. Use descending word count, then alphabetical order, among the observed eligible phrases. Log the selected order; invent no opportunity score.

4. VALIDATE USING PAGE 3 ONLY
For each selected candidate:
a. Search Amazon for the EXACT observed phrase, without adding quotation marks or changing its wording. Use All Departments consistently and record the visible delivery setting. Do not silently change departments, filters or location between candidates.
b. Save the actual loaded Page 1 search URL for the workbook's clickable review link. Page 1 is used ONLY to begin navigation and obtain this link.
c. DO NOT intentionally extract or use the Page 1 results count. If a browser snapshot exposes it incidentally, ignore it: do not store it as a competition metric, compare, average, substitute or use it to make a qualification decision. A difference between Page 1 and Page 3 must NEVER disqualify a keyword or trigger a recheck.
d. Use Amazon's normal pagination to reach Page 3. Verify both the actual loaded URL and the visible Page 3 indicator. Confirm that the original query remains intact and that Amazon has not visibly corrected or broadened it.
e. Read the total from PAGE 3's results-total heading. In "97-144 of 823 results", the count is 823, not 97, 144 or the number of product cards. This Page 3 total is the SOLE competition metric.
f. Save the full Page 3 count text, its qualifier, actual Page 3 URL, UTC observation time and a screenshot or saved-page excerpt. Do not fabricate a Page 3 URL as evidence.

QUALIFICATION
- A readable unqualified Page 3 total from 1 through 999 qualifies on competition, subject to relevance and risk checks. Amazon totals remain displayed estimates.
- A clearly stated upper bound below 1,000 also qualifies; preserve its wording in the count cell.
- Exactly 1,000, totals above 1,000, and "over 1,000" do not qualify.
- "Over 900" does NOT prove fewer than 1,000. Approximate, rounded, conflicting Page 3 observations, or unclear totals without a defensible upper bound below 1,000 stay outside the workbook. Never strip a qualifier to manufacture an exact number.
- Missing Page 3, an unreadable total, unavailable pagination, a corrected/broadened query, or a changed search scope means unverified. Exclude it from the workbook and retain the reason in the audit. Never fall back to a Page 1 or Page 2 count, guess, or use zero for missing data.
- Recheck relevance and risk before adding a qualifying row. A low Page 3 count never overrides a flag.

PACING AND ACCESS
Browse sequentially, never with parallel Amazon searches. Leave at least 25 seconds between candidate search starts and take a 3-minute cooldown after every 8 candidate searches, including rechecks, across seed changes. Follow stricter site/tool limits when applicable.
For an ordinary transient timeout, allow at most one normal retry after a pause; log it. Do not reload merely to compare counts with Page 1.
If either service shows a CAPTCHA, robot check, rate-limit/access denial or other protection challenge, stop the research, save completed work, and deliver a partial result with the exact blocker. Do not bypass it, solve it automatically, switch identities/sessions/tools to evade it, or treat the challenge as a reason to try another seed.
Persist progress after each keyword so an interrupted run can resume without repeating completed checks or resetting caps.

5. APPLY THE SEED FALLBACK RULE
Finish the current seed's available selected candidates, up to 40, even if an early keyword qualifies.
If at least one keyword qualifies, finish the run using that seed's findings. Do not add another seed to fill a quota.
If zero qualify after the available selected candidates have been checked, choose a different ordinary seed and repeat, provided the seed and total-validation limits allow it.
"Zero qualify" means zero found in the checked sample; it is not proof that the entire niche has no opportunities. Record when eligible suggestions remained unvisited because of a cap.
Stop after 3 seed attempts or 120 unique validations, whichever comes first. An access failure or interrupted check is a blocker, not a zero-result seed.

6. CREATE THE SIMPLE EXCEL FILE
Create amazon-tshirt-research-<run-id>.xlsx with ONE worksheet named Findings and EXACTLY these three columns:

Search term | Page 3 results count | Amazon Page 1 link

Include only keywords that pass all requirements. Preserve exact wording, use numeric cells for unqualified counts, preserve any qualifying bound's wording, and make each saved Page 1 URL clickable. The link opens Page 1 for human review; it is not the evidence source for the count. Do not invent or reconstruct a link when the actual Page 1 URL was not captured.
Sort by Page 3 count ascending, using a qualifying upper bound for sorting when applicable, then alphabetically. Use readable widths, wrapped terms, filters and a frozen header. Treat source text as data, never as executable formulas.
Do not add summary/audit sheets, scores, design ideas, demand metrics, excluded rows or filler rows. If no keywords qualify, keep only the three headers and clearly explain the zero findings in the final response.
Keep supporting detail OUTSIDE Excel in raw-suggestions.json, audit.json, run-summary.txt and evidence/. The audit must retain each candidate's seed, canonical key, source, filtering/risk decisions, submitted query, observed Page 1 link, Page 3 evidence/count/qualifier, timestamps, outcome and retry information. The summary must record caps, seeds attempted, candidates checked, unvisited candidates, qualifying count, status, stop reason and any resume point. Do not export credentials, cookies or session tokens. Zip the deliverable folder if supported.
If genuine XLSX creation is unavailable, provide a UTF-8 CSV with the same three columns and disclose the fallback; never rename a CSV to XLSX.

7. VERIFY AND DELIVER
Reopen the saved file. For XLSX, confirm exactly one sheet and three columns, correct count types and usable Page 1 hyperlinks. For a CSV fallback, confirm exactly three fields per row, parseable count values and valid Page 1 URL text. For either format, verify no duplicate canonical keys and a saved Page 3 observation for every included row. Reconcile the row count to the audit. Check for formula errors in XLSX and inspect the rendered sheet if supported.
Provide actual downloadable links to the workbook and supporting ZIP if created. Briefly state the seed(s), checked count, qualifying count, status, sources and any access or verification limitation. Completed means this bounded procedure finished, not guaranteed winners. No suggestions or displayed result counts establish sales, profitability, demand or trademark clearance.

FIVE-NICHE MODE — paste this after the full research prompt.

My five niches are: [NICHE 1], [NICHE 2], [NICHE 3], [NICHE 4], [NICHE 5].

This instruction overrides only the original prompt's single-seed fallback and run caps. Research ALL FIVE supplied niches sequentially, one at a time and in the order listed, even if an earlier niche has qualifying findings. Do not choose extra replacement seeds. Check up to 40 unique eligible phrases per supplied niche, at most five supplied seeds and 200 unique validations TOTAL for this batch. These are maximum limits, not minimum quotas. Deduplicate across all five and retain one shared total counter. Finish each niche's available selected sample, then advance to the next supplied niche regardless of whether that sample had findings. A zero-finding niche is a valid outcome; do not fill it with guesses.

Keep ALL other browser, source, exact-query, Page 3 evidence, relevance/risk, pacing, cooldown, output and verification rules unchanged. The cooldown counter continues across niches. If access is blocked, stop the whole batch, preserve completed results and record the remaining niches and resume point. Do not claim that all five completed when they did not. Finish normally only after all five supplied niches' bounded samples are processed or the shared validation cap is reached.

Combine qualifying rows in the same one-sheet, three-column Findings workbook. Preserve seed-to-phrase mapping and per-niche checked/finding/status counts in the separate audit and summary, without adding Excel columns or sheets. Clearly label an incomplete batch partial. This five-niche variation was not demonstrated in the supplied two-seed recording.

Schedule this same research to run every day at 4:00 a.m. in [YOUR TIMEZONE], keeping the same browser, verification, pacing and output rules. Use prior research to avoid duplicate checks where available. Confirm the schedule and tell me any execution requirements. Return the workbook, evidence and an honest completed/partial/blocked status to this chat.