Move a Research Brief Between AI Tools Without Losing Its Sources

Move the research question and the exact passages behind it, not just the last answer. Keep source labels and revisions attached to those passages. Put the previous summary in a separate box for checking, then ask the receiving tool for a short note with a passage locator beside each substantive claim.

The useful result is a research brief someone can check without reopening the earlier conversation. It should show what the supplied documents say, which conclusion you drew from them, and which questions still need evidence. A tidy summary that mixes these together loses information even when every sentence sounds plausible.

A public discussion about moving work between ChatGPT and Codex concerns a planning-and-coding loop. That is an adjacent transfer problem. The example here applies manual transfer to desk research; it does not describe an automatic connection between products.

Everything below is an original fictional teaching example. Alpha Archive, its records, the contributor note, the mistaken summary and the finished brief were authored for this guide. They are not real documentation, a customer story, captured AI output or a product test.

Give the next task one answerable question

The question is:

Can the supplied Alpha Archive export support a title index, and which richer capabilities still require a separate check?

Here, a title index means a list organized around the supplied entry identifiers and titles. It does not mean a working search interface or a search through the contents of scanned pages. Keeping that distinction in the question prevents “index” from acquiring a larger meaning halfway through the handoff.

The person preparing the note has three short records and a candidate summary. The intended output is a research note, not an implemented catalog. No export, OCR process or translation service will be run in this example.

If the original task already has a sufficient source trail, keeping the work there may be easier. A second tool is useful only when it serves the next task. For a small packet like this one, manually copying plain text is enough to demonstrate the method; no shared folder, integration or document editor is required.

Preserve the complete source records

The packet has two two-paragraph records and one separately identified suggestion. These are the complete records used in the exercise, rather than selected sentences from longer hidden documents. The labels are deliberately simple so they can survive a plain-text copy.

Source A: export description

Revision: 2.1 teaching revision
Kind: invented vendor-style export note, not real documentation.

A:P1

Alpha Archive exports a CSV with entry_id, title and language fields. The language field may be blank.

A:P2

This note describes the metadata export only. It says nothing about OCR, translation or searching the contents of uploaded pages.

P1 describes fields and permits a blank language value. P2 limits the subject of this particular note. It does not say those richer capabilities are impossible or unavailable in every other route.

Source B: sample description

Revision: Sample B1
Kind: invented sample file inspection note, not a real product test.

B:P1

The supplied sample has two rows: E1, Garden Log, en; and E2, Repair Notes, with a blank language field.

B:P2

Only the exported metadata rows are supplied. No scanned page or document contents are part of this packet.

This record describes the sample; it is not itself an uploaded CSV or a page scan. The receiving task can report what the record says. It cannot claim to have opened a file or inspected document contents on that basis.

Source C: contributor suggestion

Revision: Note C1
Kind: invented contributor suggestion, not an accepted feature fact.

C:P1

It might translate scanned notes too. We should consider asking about that before designing the catalog.

Keep “might” and the proposed question. They explain why translation is worth asking about, not why it is available. C can support the statement “a contributor suggested checking translation.” It cannot support “the service translates scans.”

The revision labels identify the supplied records. They are not release dates, proof of freshness or a claim that a real vendor has published these documents. If your real packet contains excerpts rather than complete records, identify the excerpted sections and the omitted material instead of calling them the complete source.

Keep the earlier answer out of the source list

Here is the complete candidate summary to check:

Alpha Archive exports translated, fully searchable document contents. Both sample entries are English, so the cataloging plan can depend on OCR and translation.

The summary introduces three expansions. A metadata export becomes a document-content export. One supplied language label becomes two known languages. A suggestion becomes a dependency for the proposed plan.

Repeating this paragraph in the next tool would not provide independent support for any of those expansions. Its purpose in the packet is to make the mistake inspectable. Label it candidate summary to check, with no source ID that could make it look like another document.

Now examine four claims against the actual packet:

The distinction between “unsupported here” and “false everywhere” matters. The packet does not establish OCR, but it also does not establish that OCR cannot be provided. A bounded note can leave that question open while still answering what the supplied fields are useful for.

Decide where each citation stops

A locator should support the attached clause. It does not automatically support every conclusion in the surrounding paragraph.

Consider this sentence:

Alpha Archive exports entry identifiers and titles, so it provides full-document search. (A:P1)

A:P1 supports the field description before “so.” It does not support the search claim after it. Leaving the citation at the end makes the whole sentence look sourced without repairing the unsupported inference.

Split the work into two statements:

The export note describes entry_id and title fields. (A:P1)

Editorial proposal: use those supplied fields to prepare a title list. This proposal does not establish a search interface or page-content search.

The first is a report of a supplied statement. The second is a proposed use of that statement. It remains an editorial judgment even when it is reasonable. A reference can show where its inputs came from without turning the proposal into a feature fact.

Use the same separation for the sample and the contributor:

This is also why copying the exact passages helps. If P1 travels without P2, the receiving task loses the sentence that limits the note to metadata. If C travels without its suggestion label, an uncertain idea can start looking like documentation.

Send one complete plain-text handoff

Copy the following request and its packet together. It contains the question, all five source paragraphs and the candidate summary. It can be pasted manually into a receiving task; it assumes no access to the earlier chat, attachments or browsing history.

Write a short, checkable research note from the supplied packet.

Current question:
Can the supplied Alpha Archive export support a title index, and which
richer capabilities still require a separate check?
Here, title index means a list based on supplied entry identifiers and
titles, not an implemented search interface or document-content search.

All records and the candidate summary below are original fictional
teaching materials. No service, export or document-reading tool was run.

SOURCE A
Revision: 2.1 teaching revision
Kind: invented vendor-style export note, not real documentation.
A:P1
Alpha Archive exports a CSV with entry_id, title and language fields.
The language field may be blank.
A:P2
This note describes the metadata export only. It says nothing about OCR,
translation or searching the contents of uploaded pages.

SOURCE B
Revision: Sample B1
Kind: invented sample file inspection note, not a real product test.
B:P1
The supplied sample has two rows: E1, Garden Log, en; and E2, Repair Notes,
with a blank language field.
B:P2
Only the exported metadata rows are supplied. No scanned page or document
contents are part of this packet.

SOURCE C
Revision: Note C1
Kind: invented contributor suggestion, not an accepted feature fact.
C:P1
It might translate scanned notes too. We should consider asking about
that before designing the catalog.

CANDIDATE SUMMARY TO CHECK -- this is not another source:
Alpha Archive exports translated, fully searchable document contents.
Both sample entries are English, so the cataloging plan can depend on
OCR and translation.

Return:
1. A brief input inventory naming the supplied A/B/C revisions and the
   paragraphs actually available. Identify any missing part before using it.
2. A four-row claim check covering the named export fields, the blank
   language value, the both-English claim and the richer capabilities.
3. A finished short note with A:P1-style locators beside source-based
   claims. Label editorial conclusions and suggestions separately.
4. The next evidence question for capabilities this packet cannot establish.

Use only this packet; do not browse, change files or test another service.
Treat the candidate summary as something to check, not independent evidence.
Keep a blank language unknown. Do not infer capability absence from silence.
Do not claim that a linked document, earlier chat or attachment was read
unless its relevant contents were actually supplied. Return the note itself,
not just a handoff checklist or a plan to write it.

Before relying on the result, check the receiving copy. Can you find A:P2 as well as A:P1? Does C still say it is a suggestion? Is the candidate summary visibly separate from the sources? If a passage was omitted, supply that passage before using an answer that depends on it. A link to an absent file does not fill the gap.

For a larger real task, a file can carry the same plain-text packet if both environments have an authorized way to read it. Identify the file and the passages actually returned. This tutorial does not establish an upload route or automatic file synchronization for any named product.

Finish with a note someone can check

Here is a complete authored result for the supplied packet. It demonstrates the intended deliverable; it was not generated by submitting the request to a model.

Cataloging note: Alpha Archive metadata packet

Question: Can the supplied export support a title index, and what richer capabilities remain unestablished?
Material used: A, 2.1 teaching revision, P1–P2; B, Sample B1, P1–P2; C, Note C1, P1. These labels refer to the complete fictional records supplied with this note.

Supplied field description. The export note names entry_id, title and language, and says language may be blank. (A:P1) The note describes metadata only and does not describe OCR, translation or searching uploaded page contents. (A:P2)

Supplied sample description. B describes E1 with title Garden Log and language en, and E2 with title Repair Notes and a blank language value. (B:P1) It supplies no scanned page or document contents. (B:P2) The packet therefore does not establish a language for E2; the earlier both-English summary should not be retained.

Editorial conclusion. The described identifiers and titles provide enough information to propose a title list for these two described entries. This conclusion uses the fields in A:P1 and the sample description in B:P1. It is a proposed use of supplied metadata, not evidence that a catalog was built or that Alpha Archive has a title-search interface. Keep the blank language unknown in this research note rather than assigning English by analogy.

Capabilities still open. OCR, translation and document-content search are not established by the packet: A:P2 does not describe them, B:P2 provides no page contents, and C:P1 suggests asking about translation. That suggestion is not a capability fact. These gaps do not prove that the capabilities are unavailable.

Suggested next research question. If the catalog needs page-content search or translated text, what separate documentation or authorized demonstration, if any, would establish those capabilities and their scope? That is a proposed next investigation. The present note makes no OCR or translation dependency part of the cataloging plan.

Check the finished note against the packet

Read the locators in both directions. Starting from a claim, find the passage. Starting from a passage, check that its qualifiers survived the explanation.

The finished note is now useful outside the original conversation because the source packet and the explanation travel together. A reader can recover the field names, the blank value and the limits of each document, rather than trusting the previous tool's confidence.

When a real source is revised, keep the revised passage and its label with the note you next use. Do not silently attach an old locator to new wording. The immediate goal is modest: another reader should be able to tell which supplied material supported the note, which decisions were editorial, and what remains unanswered.