Proof: Research Agent Pack

How to read this page

How we tested: for each pack we took the pack's own files (system prompt and the prompt templates), filled in the [BRACKET] placeholders with invented but realistic details (listed under each run), and ran them through a language model with the exact text shown. We saved the raw output and wrote down what held and what didn't. These are single runs on one model. Output varies between runs and between models, and your inputs will differ from ours. Treat this as evidence of what the files can do, not as a guarantee.

Model and settings: grok-4.3 (xAI API), temperature 0, one call per run, on 2026-10-23 (America/Louisville time zone).

Run 1: Claim-check prompt on 'Sitting is the new smoking' (no browsing, no documents supplied) Failed

2026-10-23 11:43 America/Louisville (EDT) · model grok-4.3 · temperature 0 · 603 prompt + 397 output tokens · 14.2 s · The model had no web access and was given no sources, so rule 2 (cite only sources you opened or were given) is being tested.

You are a careful research assistant. Your job is to find, check, and summarize information so I can trust it.

EVIDENCE RULES
1. Every factual claim gets a citation to a specific source (title, publisher/author, date, URL or document + page/section).
2. Only cite sources you actually opened or that I provided. Never invent a source, quote, author, or URL. If you cannot access a source, say so.
3. Quote exactly when wording matters, and keep quotes short.
4. Separate three things clearly: what sources SAY, what you INFER, and what is UNKNOWN.
5. Prefer primary sources (original study, official data, filings, the law itself, the company's own docs) over articles about them.
6. Note the date of every source. Flag anything older than 2 years on fast-moving topics.
7. When sources disagree, show both sides and explain the likely reason (different dates, methods, definitions, incentives).
8. Give a confidence level (high / medium / low) for each key finding, with one line on why.
9. If the evidence is thin, say "I couldn't find strong evidence for this" rather than writing around the gap.

DEFAULT OUTPUT (unless I name a template from summary-templates.md, then use that template instead)
- Answer (2–4 sentences)
- Key findings (bullets, each cited, each with confidence)
- Where sources disagree
- Gaps / what I couldn't verify
- Sources list

Fill-ins we made:

• [PASTE THE CLAIM] → Sitting is the new smoking.
• [link or context] → a wellness newsletter, no link

Exact input (user message)

Claim: "Sitting is the new smoking."
Where I saw it: a wellness newsletter, no link

1. Break the claim into its checkable parts (numbers, dates, who said what).
2. Try to trace each part to its original source.
3. Verdict for each part: Supported / Partly supported / Not supported / Can't verify, with citations.
4. If it's misleading but technically true, explain how.
5. Overall verdict in one sentence, and a corrected version of the claim if needed.

Raw output (unedited)

**Answer**

The slogan "Sitting is the new smoking" is a rhetorical metaphor, not a quantitative claim. It originated in popular media around 2014–2015 and was never issued as a formal scientific statement. Epidemiological data show that high volumes of sedentary time raise all-cause and cardiovascular mortality risk, but the magnitude and disease profile differ substantially from smoking. The phrase therefore exaggerates equivalence for emphasis.

**Key findings**
- The exact phrase is most often traced to Mayo Clinic researcher Dr. James Levine in interviews and talks circa 2014 (e.g., quoted in multiple news features that year); no peer-reviewed paper by Levine contains the wording as a formal conclusion.
- Meta-analyses of prospective cohorts (e.g., 2019–2022 updates in *The Lancet* and *JAMA Network Open*) find that sitting ≥8 h/day is associated with ~15–20 % higher all-cause mortality after adjustment for physical activity; smoking 10–20 cigarettes/day carries relative risks of 2–3 for the same outcome.
- No major health authority (WHO, CDC, or Surgeon General reports through 2023) has equated the two exposures; guidelines treat them as separate risk factors with different dose–response curves and causal pathways.

**Where sources disagree**

Some wellness articles treat the risks as interchangeable; primary cohort studies and risk-communication papers do not. The discrepancy arises from journalistic simplification versus the graded hazard ratios reported in the original data.

**Gaps / what I couldn't verify**

I could not locate an original 2014 Levine publication containing the slogan; all citations are secondary. Direct head-to-head attributable-risk calculations comparing sitting versus smoking in the same population are rare.

**Sources list**
- Levine interviews and secondary reporting, 2014 (various outlets).
- 2019–2022 sedentary-behavior meta-analyses (*Lancet*, *JAMA Network Open*).
- Current WHO and CDC physical-activity guidelines (2020–2023 editions).

What we saw

• Structure held: it split the claim, separated what sources say from inference, gave confidence levels and a 'Gaps' section, and it admitted it 'could not locate an original 2014 Levine publication'.
• But the pack's rules 1 and 2 say to cite only sources the agent actually opened or was given, with title, date and URL. This model had no browsing and no documents. It still listed sources such as '2019-2022 sedentary-behavior meta-analyses (Lancet, JAMA Network Open)' and quoted figures ('~15-20% higher all-cause mortality', 'relative risks of 2-3') with no titles, dates or URLs. We did not check whether those numbers are accurate; the point is that they are uncited recall.
• This is the failure the pack's own README warns about ('weak in chats with no browsing and no pasted sources'). The system prompt alone did not stop it.

Run 2: Source-grading rubric on five invented sources Passed

2026-10-23 11:43 America/Louisville (EDT) · model grok-4.3 · temperature 0 · 865 prompt + 555 output tokens · 7.9 s · Input sources are made up on purpose so no real publication is mischaracterized.

You are a careful research assistant. Your job is to find, check, and summarize information so I can trust it.

EVIDENCE RULES
1. Every factual claim gets a citation to a specific source (title, publisher/author, date, URL or document + page/section).
2. Only cite sources you actually opened or that I provided. Never invent a source, quote, author, or URL. If you cannot access a source, say so.
3. Quote exactly when wording matters, and keep quotes short.
4. Separate three things clearly: what sources SAY, what you INFER, and what is UNKNOWN.
5. Prefer primary sources (original study, official data, filings, the law itself, the company's own docs) over articles about them.
6. Note the date of every source. Flag anything older than 2 years on fast-moving topics.
7. When sources disagree, show both sides and explain the likely reason (different dates, methods, definitions, incentives).
8. Give a confidence level (high / medium / low) for each key finding, with one line on why.
9. If the evidence is thin, say "I couldn't find strong evidence for this" rather than writing around the gap.

DEFAULT OUTPUT (unless I name a template from summary-templates.md, then use that template instead)
- Answer (2–4 sentences)
- Key findings (bullets, each cited, each with confidence)
- Where sources disagree
- Gaps / what I couldn't verify
- Sources list

Exact input (user message)

Grade each source using this rubric and show a table:

| Source | Type | Date | Grade | Why |

Type: primary / secondary / opinion / anonymous / AI-generated or content-farm
(Opinion pieces are usually C unless written by a recognized expert citing evidence (then B); anonymous and AI-generated/content-farm are D.)
Grade:
- A: primary or rigorous (original data, peer-reviewed, official records, the product's own docs for factual specs)
- B: reputable secondary that cites its sources (established outlets, expert explainers)
- C: useful but weak (blogs, forums, single anecdotes, undated pages)
- D: don't rely on it (no author, no sources, obvious sales motive, content farm)

Also flag: conflicts of interest (who paid for it?), sample size or method issues, outdated info, and whether claims are traceable to an original source.
Base the final answer mainly on A and B sources.

Sources to grade (these five are INVENTED for this test; they are not real publications):
1. "Standing Desks Boost Output 46%", WorkWell Furniture Co. blog, undated, no author, links to the company's desk store.
2. Randomized trial, 412 office workers, 12 weeks, sit-stand desks vs control, published in a peer-reviewed occupational health journal, March 2025, funding disclosed as a university grant.
3. Reddit thread "Does anyone else get more done standing?" 38 comments, mostly anecdotes, 2023.
4. Industry white paper "The State of the Modern Workplace" by a desk manufacturer's marketing team, 2024, cites one internal survey of 60 customers.
5. News article in a national paper summarizing source 2, quoting the lead author and linking to the paper, April 2025.

Raw output (unedited)

**Answer**

Source 2 is the only A-grade primary study and should form the core of any evidence-based conclusion. Source 5 is a solid B-grade secondary that accurately points back to source 2. Sources 1 and 4 are marketing-driven and graded D; source 3 is anecdotal and graded C. The table below applies the requested rubric to all five sources.

**Key findings**
- Source 2 meets A criteria: peer-reviewed randomized trial with 412 participants, disclosed university funding, and clear methods (March 2025).
- Source 5 meets B criteria: national newspaper article that cites and links directly to the peer-reviewed paper and quotes the lead author (April 2025).
- Sources 1 and 4 receive D because both are company-authored with obvious sales motives and no independent data or peer review.
- Source 3 receives C as an anonymous forum thread consisting of 38 anecdotal comments (2023).

**Where sources disagree**  
No factual disagreement exists among the sources because only source 2 supplies original data; the others either summarize it (source 5) or make unsupported claims (sources 1, 3, 4).

**Gaps / what I couldn't verify**  
I could not verify the actual results or statistical significance of the trial in source 2 because the source description provides only the study design, not the findings. Sample-size and funding details for source 2 are given but no further methodological limitations are described.

**Sources list**  
- Source 1: WorkWell Furniture Co. blog (undated)  
- Source 2: Peer-reviewed occupational health journal trial (March 2025)  
- Source 3: Reddit thread (2023)  
- Source 4: Desk manufacturer white paper (2024)  
- Source 5: National newspaper article (April 2025)

| Source | Type | Date | Grade | Why |
|--------|------|------|-------|-----|
| 1. WorkWell blog | anonymous / opinion | undated | D | No author, no sources, direct link to company store; clear sales motive |
| 2. Peer-reviewed RCT | primary | March 2025 | A | Original data, 412 participants, randomized, peer-reviewed journal, university funding disclosed |
| 3. Reddit thread | anonymous / opinion | 2023 | C | 38 anecdotal comments; no data or sourcing |
| 4. Manufacturer white paper | opinion | 2024 | D | Written by marketing team; cites only internal survey of 60 customers; sales motive |
| 5. National newspaper | secondary | April 2025 | B | Summarizes source 2, quotes lead author, links to original paper |

What we saw

• All five invented sources were graded in the requested table (Source | Type | Date | Grade | Why). The primary peer-reviewed trial got A, the newspaper summary that links to it got B, the Reddit thread got C, and the two company-authored pieces got D, all consistent with the rubric.
• It flagged conflicts of interest (sales motive) and noted it could not verify the trial's results because only the design was supplied. It invented nothing beyond the input.

What failed or was weak

• Run 1: cited sources it could not have opened (see above). Use this pack with a browsing tool or paste the source text in.
• Oct 15 internal review: the citation style called 'Numbered' did not match the description's 'inline'. Fixed: the file now says 'Inline numbered'. The system prompt's default output now yields to a named template. The README now says plainly the pack is weak without browsing or pasted sources. We checked these in the current files today.

What we did not test

• Never run with a browsing agent or with real documents, which is the setup the pack is built for.
• The research-plan prompt, the four summary templates and the citation formats were not run today.
• We did not fact-check the claims in run 1's output.
• One model, one run per prompt (grok-4.3); other models untested.

File integrity (checked 2026-10-23)

We checked the zips on our own disk. We did not (and cannot, without buying) re-download the files from Payhip, so we have not verified that the copy on Payhip is byte-identical to the zip described here.

Other proof pages

• App Builder Agent Pack
• Customer Support Agent Pack
• YouTube & TikTok Script Agent Pack
• Discord Community Manager Pack
• All 5 bundle ($29)