/source-sifter

---
name: source-sifter
description: >-
  Read a pile of sources (reports, articles, PDFs, filings, transcripts,
  web results) and return one ranked brief: what the evidence agrees on,
  where sources contradict each other, how strong each claim is, and what
  is still unknown, with every claim traced to its source and date. Use
  this skill whenever the user uploads or links several documents and asks
  to "summarise these", "what do these say", "synthesise", "research this
  topic", "compare these reports", "literature review", "what's the
  consensus", or is preparing a recommendation, investment thesis, board
  paper or article from multiple sources. Do NOT use for a single short
  document (just summarise it), for fact-checking one draft (use a claim
  check), or when the user wants opinion without sources.
---

# Source sifter

Ten sources do not give you ten times the truth. Usually they give you
three original findings, repeated, rounded and reworded until they look
like consensus. A summary that treats every source as equal ends up
averaging a primary study with a blog post that misread it.

The job here is not to compress. It is to sort: which claims rest on
evidence, which rest on repetition, where serious sources disagree, and
what nobody has measured yet. A decision-maker can work with "we don't
know" much better than with a confident summary that hides a split.

## Workflow

### 1. Log the sources

Make a table before reading for content:

| # | Source | Type | Date | Who made it | Primary or secondary |
|---|---|---|---|---|---|

Type: study, official statistics, company filing, industry report, news,
opinion, transcript. Primary means the source produced the data or was in
the room. Flag anything older than the question allows (a 2021 market size
for a 2026 decision).

### 2. Extract claims, not paragraphs

For each source, pull the specific claims that bear on the user's
question, with the number, the scope (who, where, when) and the page. Keep
wording precise: "associated with" and "causes" are different claims.

### 3. Cluster and trace

Group claims that say the same thing. For each cluster, trace it back: do
the sources cite each other or the same origin? Five articles quoting one
survey count as one source, and the brief must say so.

### 4. Grade each finding

- **STRONG:** two or more independent primary sources agree, recent,
  method visible.
- **MODERATE:** one good primary source, or several secondary ones that
  trace back to different origins.
- **WEAK:** secondary only, single origin, old, or method unclear.
- **CONTESTED:** credible sources disagree. Show both sides and the
  likely reason (different period, sample, definition).

### 5. Write the brief

```
## [Question], from N sources, as of [date]

**Bottom line:** [2-3 sentences: what the evidence supports, how strongly,
and the biggest open question.]

**Findings, strongest first**
1. [Finding]: STRONG. Sources: [#2 p.14, #5 table 3]. [One line on why it holds.]
2. ...

**Where sources disagree**
- [Topic]: #3 says X (2025, survey of 2,000), #7 says Y (2023, 300 firms).
  Likely reason: [ ].

**What nobody here measured**
- [Gap that matters for the decision]

**Source quality notes:** [duplicates, outdated items, conflicts of interest]
```

Never pad a finding to fill the list.

## Judgment rules

**Surface disagreement, never average it.** If one source says the market
grows 4% and another 12%, the answer is not 8%. Show both and why.

**Facts, inferences and opinions stay labelled.** "The filing reports X"
is a fact. "This suggests Y" is your inference: say so.

**Count origins, not citations.** A widely repeated figure with no
findable origin is WEAK, however many sites carry it.

**Interests matter.** A vendor's report on the market for its own product
is still evidence, but the brief notes who paid for it.

**No invented sources.** Every claim points to a source the user gave or
one you actually retrieved. If you cannot find support, say "not found".

## Example

Seven sources on "do four-day weeks cut productivity?". Three news
articles, one consultancy blog and one LinkedIn post all trace back to the
same UK pilot. The brief counts that as one MODERATE source, lists a
separate Icelandic public-sector trial as the second, notes that both
measured self-reported productivity, and puts "no study here measured
output in finance roles" under gaps. The user now knows the evidence is
thinner than the headlines suggested.