media_literacy v1.0.0

name: "media_literacy" description: "Verify a news story, photo, or viral claim in minutes: lateral reading, reverse image search, and headline-vs-body checks. Trigger when asked to check whether a story, image, headline, or claim is trustworthy, or when you need to source a claim before repeating it." metadata: { "includeInPrompt": true }

Media Literacy

Purpose

Give an agent a fast, repeatable way to answer one question: can I trust this thing I'm looking at? Three drills cover the three things that actually mislead people: a story with no solid sourcing, a manipulated or recycled image, and a headline that outruns its own article. Each drill takes about five minutes. Details live in references/drills.md; the printable pocket version is references/quick_reference.md.

Workflow

Pick the drill that matches the material. You can run more than one on the same item.

Drill 1: Source the story (5 minutes). Use SIFT: Stop (pause before believing or sharing), Investigate the source (leave the page, search who the publisher/author is, read what others say about them), Find better coverage (search the claim's core; see if independent outlets like Reuters or AP confirm it), Trace to the original (follow links to the underlying study, document, or quote; check the date). If a fact-checker already did the work (Snopes, PolitiFact, FactCheck.org, Full Fact), check for previous work first and don't redo it.

Drill 2: Check the image. Run a reverse image search on two engines before anything else: TinEye (earliest-instance tracking) plus Google Lens or Bing Visual Search (visual similarity). Ask: does the earliest appearance predate this story? Does it show a different place, event, or year? Then eyeball for manipulation: AI tells (hands, teeth, background edges, mismatched symmetry), inconsistent lighting or shadows, watermarks, and stripped EXIF on a photo that claims a specific time and place. A recycled image from an older event is the most common trick, and reverse search catches it. Absence of artifacts does not prove authenticity.

Drill 3: Read the headline against the article. Read the headline alone first and write down exactly what a reasonable reader would take from it. Then read the body and check each claim: supported, partially supported, unsupported, or contradicted. Flag buried qualifications (the key hedge sitting in paragraph 8+), tense drift (headline asserts what the body says was only alleged), and actor switching (headline names one party, body reveals another). A headline that misleads while the body stays honest is a mismatch, not a lie, and it should be reported as one.

Output Contract

Return a verdict with this shape:

• Rating: one of Verified, Likely real, Unverifiable, Misleading, False.
• One-line summary of what the material claims.
• Evidence: the strongest 2-4 findings, each tied to a drill and a source. Link or name every source.
• Gaps: what you could not check (paywall, truncated body, no reverse-search history).
• Confidence: high / medium / low, with one sentence saying why.

A template with annotation guidance is in references/verdict_template.md.

Operating Rules

1. Every factual claim in a verdict must come from a source you actually looked at. Never invent numbers, dates, quotes, or events. Unverifiable means you say so and drop the rating a level.
2. Lateral reading beats vertical reading. Do not decide a source is trustworthy by reading the source's own about page. Leave the page and research it elsewhere.
3. One engine is a lead, not a confirmation. Reverse-search in two places; corroborate claims across independent outlets before calling anything Verified.
4. Rate the material, not the outlet. A solid outlet can run a bad headline; a bad outlet can run a true photo. Keep drill results separate from prior reputation.
5. If you catch yourself feeling certain fast, slow down. Strong emotion (outrage, vindication, alarm) is a signal to re-run Drill 1, not a finding.
6. Never share or republish the material while the rating is Unverifiable or worse. Say what is missing and what would change the rating.