Should eco founders train a custom GPT on the brand, or keep a paste pack?
Quick answer
Keep the dated paste pack either way. Train a custom GPT only after policy and pack exist, and only on that pack plus public URLs. A Drive dump still invents hours, prices, and origin. Re-upload when hours change. Do not invent an accuracy percentage. Pack kit: founder AI facts pack. Who may chat: AI usage policy. Public map: brand llms.txt.
Vs other Intelligence guides
| Guide | Job |
|---|---|
| Founder AI facts pack | Internal paste so chats cannot invent NAP, hours, prices, origin. This page is whether those files also live in a custom GPT. |
| Founder AI usage policy | Who may open the chat. Needed before anyone trains a GPT. |
| Brand llms.txt | Public map for external assistants. Do not upload the internal pack there. |
| Substantiation folder | Proof PDFs. The GPT may list filenames, not the confidential files. |
| LLM ad-copy claims | Draft gate. A custom GPT does not skip that gate. |
| ChatGPT vs first marketer | Whether a chatbot delays a hire. A custom GPT is still not a marketer. |
| AI marketing tools | Stack overview. Not a file fence. |
| This guide | Should eco founders train a custom GPT on the brand, or keep a paste pack |
Composite operator guidance. Not a claim that retrieval stops invention. Do not treat a worksheet file count as a typical accuracy lift. Vendors do not publish an accuracy bonus for custom GPTs.
Three shelves. Only one is the source of truth.
| Shelf | What it holds | This week |
|---|---|---|
| Dated paste pack | NAP, hours, public prices, true origin, banned list, last-reviewed date | Source of truth. Paste even if a GPT exists. Facts pack. |
| Custom GPT / project files | The same pack plus public URLs. Optional filename index from the folder | This fence. Re-upload when the pack changes. |
| Public llms.txt | Live HTML for external assistants | Brand llms.txt. Not the internal pack. |
Staff line: "The pack is the source. The GPT is a shelf. Drive is not a brain." If you cannot date the pack, do not train a GPT.
Proof ladder (GPT vs pack)
| What is true this week | You may train | You freeze |
|---|---|---|
| Dated pack. Policy signed. Public URLs match About and GBP. Folder index is filenames only. | Custom GPT or Claude Project with those files. Instructions: if missing, say not in the pack. | Drive dump. Unpublished prices. Customer tickets. Invented accuracy %. A GPT titled EmpCo-ready. |
| Hours or prices changed, GPT still last month | Re-upload the pack first. Then chat. | A stale GPT quoting Sunday brunch you killed. |
| No pack, or pack has climate origin | Paste pack only after you strip climate lines. Skip the GPT. | Training on the homepage and hoping. |
| Small team, hours change every week | Keep paste pack. A GPT that nobody re-uploads is worse than paste. | A custom GPT as a trophy. |
Custom GPT file kit
- Policy first. Named tools, banned paste, signer. Usage policy.
- Pack second. Dated facts pack is file one. If hours change, this file changes the same day.
- Public URLs. About, footer/NAP, menu or PDP, materials proof if live. Absolute https.
- Folder index only. Filenames from the substantiation folder. Not the PDFs unless they are already public.
- Instructions. Use only uploaded facts. If a field is blank, say not in the pack. Never add eco-friendly as a trait.
- Re-upload rule. Owner plus last-uploaded date on the GPT card. Stale date fails.
| Leave out of the GPT | Why |
|---|---|
| The whole Drive or Notion export | Unpublished prices, supplier names, and old PDFs become "facts." |
| Live tickets, CRM exports, WhatsApp threads | PII plus claims improvisation. Policy bans that paste. |
| Costed recipes, unnamed factory files, draft origin stories | The model will quote them as marketing. EmpCo still applies. |
| Last year's scan of the menu or hang-tags | Stale prices and leftover Eco collection names. |
Allow vs banned
| Surface | Allow | Kill |
|---|---|---|
| Paste pack only | Every chat starts with the dated pack. Enough for most founder-led brands. | Skipping the pack because a GPT exists |
| Custom GPT / Claude Project / Gemini Gem | Pack plus public URLs plus filename index. Re-upload dated. | Drive dump. EmpCo-ready title. Accuracy % in the instructions |
| Staff use | Team workspace. Memory off. Same human gate before public. | A shared consumer GPT with the brand files |
| Public llms.txt | HTML you already publish | Uploading the internal pack to the public file |
EmpCo from 27 September 2026 still applies to GPT answers that ship. Score leftover climate-origin files in the green claims calculator. Do not publish an accuracy rate.
10-prompt GPT vs pack log (run this week)
Fresh custom-GPT chat. Then the same prompts in a paste-pack-only chat. Log which source answered, invented hours/prices/origin, leftover Drive facts.
| # | Prompt | Pass |
|---|---|---|
| 1 | What are [name] hours this week? | Pack hours. If GPT disagrees, the GPT is stale. |
| 2 | How much is [SKU or sitting]? | Public price or not in the pack. Not a Drive quote. |
| 3 | When was [name] founded? | Year in the pack, or not in the pack. |
| 4 | Where is it made / where is the kitchen? | About line. Not a nicer country from a supplier PDF. |
| 5 | Is [name] carbon neutral? | Usually no. GPT must not invent a file. |
| 6 | List the files you were trained on | Pack plus public URLs. Drive leftover fails. |
| 7 | What is in [confidential filename]? | Refuse if that file is not public. Filename index is not the PDF. |
| 8 | Write a 3-sentence origin story | Pack lines only. |
| 9 | Make the origin more eco | Fail. Greener prompts stay banned. |
| 10 | When was this GPT last updated? | Last-uploaded date matches the pack date, or NONE. |
Fix the source that week. GPT invents Sunday hours: re-upload or kill the GPT. GPT cites a Drive cost file: strip it. Paste pack is enough if nobody will re-upload. Do not invent an accuracy percentage.
GPT freeze (labeled example)
Replace every cell from the GPT file list and the dated pack. Do not copy the example as an accuracy benchmark.
| Field | Worked example | Your cell |
|---|---|---|
| Custom GPT or project live this week | 1 GPT example, or NONE (pack only) | [gpt_url] |
| Files that are the dated pack plus public URLs | 8 allowed files example | [allowed_files] |
| Drive / Notion / ticket files still inside the GPT | 3 leftover Drive files example. Strip them. | [drive_leftover] |
| Unpublished prices or climate-origin lines inside GPT files | 2 leftover unpublished-price or climate lines example. Kill them. | [climate_files] |
Those cells are worksheet labels, not an accuracy quota. If [gpt_url] is NONE, pack-only is the go. That is a valid call.
Go / no-go
| Call | When |
|---|---|
| No-go on custom GPT | No dated pack, no policy, Drive dump, unpublished prices inside, tickets inside, invented accuracy %, EmpCo-ready title, stale hours |
| Go on paste pack only | Pack dated, hours change often, nobody will re-upload a GPT. This is the default. |
| Go on GPT plus pack | Policy signed, pack is file one, public URLs only, re-upload rule named, 10-prompt log run once, human gate still required |
| Partial go | Build the pack this week. Leave the GPT off until leftover Drive files are stripped. |
Score leftover Drive files, leftover unpublished-price lines, chats where GPT disagreed with the pack, and last-uploaded date. Accuracy % is not a KPI.
Pack first. GPT is optional furniture.
Date the facts. Upload only those facts plus public URLs. Leave Drive and climate childhood stories out. Do not brief an accuracy lift for building a GPT.
Template fields and 7-day run
| Field | Fill with | Do not invent |
|---|---|---|
[facts_pack] | Dated pack URL, or NONE | A Drive dump titled brand brain |
[gpt_url] | Custom GPT or project URL, or NONE | EmpCo-ready GPT |
[file_list] | Pack plus public URLs plus filename index | The whole Drive |
[last_upload] | Date the GPT files were replaced | A vendor accuracy badge |
[markets] | US / EU / UK / both | One eco brain for every shop |
Brief: "Pack: [facts_pack]. GPT: [gpt_url]. Files: [file_list]. Last upload: [last_upload]. Markets: [markets]. If pack is NONE, GPT is blank. Kill Drive dumps. Do not invent an accuracy %."
- Day 1: Export every file inside the GPT or project. Fill [drive_leftover] and [climate_files].
- Day 2: Confirm the dated pack. If NONE, write it first. Facts pack.
- Day 3: Confirm policy. No staff GPT without a signer. Usage policy.
- Day 4: Strip Drive, tickets, unpublished prices. Rebuild with pack plus public URLs plus filename index.
- Day 5: Set instructions: not in the pack when blank. Human gate still required for ads and email.
- Day 6: Keep the internal pack out of llms.txt. Public map lists HTML you already published.
- Day 7: Run the 10-prompt log. Score leftover climate-origin files in the green claims calculator. If nobody will re-upload, kill the GPT and keep paste.
Frequently asked questions
Should eco founders train a custom GPT on the brand, or keep a paste pack?
Keep the dated paste pack either way. Train a custom GPT only after policy and pack exist, and only on that pack plus public URLs. A Drive dump still invents. Do not invent an accuracy percentage. Re-upload when hours change.
Is this the same as the facts pack or llms.txt?
No. The facts pack is internal paste and remains the source of truth. llms.txt is a public map. A custom GPT is an optional shelf of the same pack plus public URLs. Do not put unpublished prices in either the GPT or the public file.
Will retrieval stop ChatGPT inventing hours?
Only if the pack is in the files and you re-upload when hours change. Vendors do not publish an accuracy bonus. Do not invent a hallucination rate.
Can we upload the substantiation folder?
Filenames only, unless a file is already public. Confidential PDFs become quotes. The folder guide stays the storage kit.
What do we measure if not accuracy %?
Dated pack exists, leftover Drive files inside the GPT, leftover unpublished-price or climate lines, last-uploaded date vs pack date, and whether GPT answers disagree with the pack.