How to Build a Source Pack AI Can Use Without Losing the Evidence
September 16, 2026
Most AI-assisted articles fail verification not because the research was bad, but because the research was never organized in a way that survives contact with drafting. Someone reads six sources, has the key facts in their head, writes a prompt describing what they learned, and generates a draft. Three days later, an editor asks, “Where does this number come from?" and the honest answer is that it is somewhere in one of six browser tabs that got closed.
A source pack fixes this by giving research a fixed, reusable shape before drafting starts, one that survives being handed to an AI tool, an editor, and a client review, without losing the trail back to the original evidence.
What is a source pack?
A source pack is a table, not a paragraph. Each row is one claim you intend to use, paired with exactly where that claim came from. The draft can be generated, revised, and reviewed any number of times. The source pack does not change unless the research itself changes.
The schema has six columns:
| Column | What it holds |
|---|---|
| Claim | The specific fact, statistic, or statement you intend to make |
| Passage | The exact sentence or data point from the source, quoted, not paraphrased |
| Source URL | A link that currently loads and shows that passage |
| Date | The publication date of the source, or the date you accessed it if the source has no date |
| Scope | What the claim can and cannot be generalized to (a survey of 200 US freelancers is not "professionals," it is "US freelancers surveyed") |
| Unresolved | Anything about the claim you were not able to confirm, left explicit rather than smoothed over |
Figure 1. The six columns of a source pack, and what each one is meant to hold.
Six columns are enough for almost any article. Adding more fields makes the pack harder to fill in consistently, and a source pack that is too tedious to complete gets skipped under deadline pressure, which defeats the purpose.
A completed example
Here is a real source pack for a hypothetical article about remote work adoption, built to show what "done" looks like rather than to make a claim about the topic itself.
| Claim | Passage | Source URL | Date | Scope | Unresolved |
|---|---|---|---|---|---|
| Hybrid work has become the dominant model at large US employers | "Survey respondents at companies with 1,000+ employees reported hybrid arrangements at a higher rate than fully remote or fully in-office." | (link to the specific report page) | 2026 | US employers with 1,000+ staff surveyed in this specific report; not a global claim | Exact percentage was rounded in the summary page; the full PDF has the precise figure if needed |
| Fully remote roles are more common in technology than in other sectors | "Technology sector respondents reported fully remote arrangements at roughly twice the rate of the overall sample." | (link to the specific report page) | 2026 | This report's sample only; other surveys may define "technology sector" differently | None |
| Return-to-office mandates increased in early 2026 | (a news article describing several named companies' announcements) | (link) | 2026 | Describes named companies only, not an industry-wide statistic | This is anecdotal evidence of a trend, not a measured rate; the article should say so explicitly rather than implying a broader statistic |
Figure 2. A completed source pack for a hypothetical article, including one row marked as anecdotal rather than measured.
Notice the third row. Not every claim needs to be a hard statistic. A well-labeled anecdotal observation, clearly marked as such in the "Unresolved" column, is honest and usable.
What is not usable is the same claim written as if it were a measured industry rate, because the source pack forces you to notice the difference before it reaches the draft.
An unsupported claim, and what to do with it
Say your research turns up a compelling line you want to use: "most companies that tried mandatory return-to-office saw an increase in resignations." You cannot find a source that actually measured this across "most companies." You find two or three individual company examples and a lot of commentary asserting the pattern is common.
The source pack forces a decision at this point, rather than letting the claim slide into the draft on the strength of how often you have seen it repeated elsewhere. Your options, in order of preference:
- Narrow the claim to what you can support: "At least three large employers that enforced return-to-office mandates in 2026 reported resignation spikes shortly after, including [named examples with sources]." This is defensible and still a strong point.
- Attribute it as commentary, not fact: "Several workplace analysts have argued that mandatory return-to-office policies risk higher resignation rates, though comprehensive data across companies is limited." This is honest about the limits of the evidence.
- Cut it: If neither version serves the article as well as the original overreaching claim did, that is a sign the original claim was doing more work than the evidence supports.
What should never happen is the claim reaching the final draft in its original, unsupported form because it was more compelling than the versions you could actually stand behind.
Handing the pack to an AI drafting tool
Once the pack is complete, it becomes the input for drafting, not a reference you check after the fact. The instruction to the model should include the full table and an explicit constraint:
Prompt: "Draft an article section using only the claims in this source pack. For each claim you use, keep it worded consistently with the 'Scope' column, do not generalize beyond it. Do not introduce any statistic, study, or named source that does not appear in this table."
This inverts the usual failure mode. Instead of the model generating claims and you trying to verify them after the fact, the model is restricted to claims that were already verified before drafting began.
Keeping the pack attached through review
The most common way a source pack stops working is that it gets separated from the draft it belongs to; the table lives in one document, the draft moves to another, and by the time an editor is reviewing the piece, reconnecting the two takes as long as redoing the research.
Whatever system you use, the source pack and the draft it supports need to travel together, so a reviewer questioning a specific sentence can find its row in seconds, not minutes.
How Weelkit helps
Weelkit keeps research attached to the draft it supports all the way through review, so a reviewer questioning a specific line finds the source next to it in the same workspace, rather than reconstructing a research trail from memory or a closed browser tab.
Quick reference: the six-column template
Copy this structure directly into a spreadsheet or document to start your own pack.
| Claim | Passage | Source URL | Date | Scope | Unresolved |
|---|---|---|---|---|---|
| (blank) | (blank) | (blank) | (blank) | (blank) | (blank) |
Figure 3. The blank six-column template, ready to copy into a new source pack.
Frequently Asked Questions
No. It is a table, not a piece of software, so a spreadsheet or a plain document with the six columns works fine. What matters is that the pack stays attached to the draft it supports, not which tool holds it.
See how Weelkit fits your team.
Request access. We will get in touch to discuss your workflow and help connect your CMS and analytics.