Index of every page on this site: llms.txt
All skills
wshobson avatar

/social-evidence-research

@4220308
by Seth Hobsonwshobson/agents43k stars
4,518
Run it yourself

Research public discussion on social platforms and keep every report claim tied to the source material that supports it. Use when researching or comparing how a product, event, or topic is discussed across social platforms or supplied social exports, when turning search results into a cited report, or when auditing a social research report for discovery-only evidence, double-counted reposts, misattributed comments, claims that go beyond the sample, or hidden access gaps.

Use this Skill: https://skilld.dev/gh/wshobson/agents/social-evidence-research

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

referencesacceptance-examples.md

≈2.2k tokens on demand. Your agent reads this file only when SKILL.md points to it.

Acceptance examples

Each example gives compact input records and the result the method must produce. A reviewer can check them by reading the rules in SKILL.md, evidence-record.md, and report-template.md, with no accounts, browsing, or paid services. Records show only the fields the example needs. All handles and URLs are made up.

Summary

Input Expected result
One inspected caption, three search cards, and an unwatched attached video. Only the caption supports claims about the post's text. The cards remain discovery-only, and the video's contents remain unknown.
An original post, two reposts, and a disagreeing comment. The repeated original counts once as support. The comment is attributed to its commenter and does not change what the post author said.
Three supportive comments among twelve inspected comments selected by a stated method. The report can say "3 of 12 inspected comments expressed support," with citations and sample limits. It cannot infer that 25% of the audience supports the topic.
A blocked source and an empty search result. Both appear in the access and collection limits. The report makes no claim that the topic has no discussion.
A transcript read, with the video unwatched. Claims cover the transcript text only. What the video shows stays unknown.
An export with a capture date and origin. Claims cite the export as a capture from that date, never as a live check.
A repost with added commentary. The original counts once. The added commentary is a separate statement by the reposting author.
A source stating a fact. The report says the source stated it. It doesn't present the fact as true.
A harness without browsing, given only URLs. Every URL is discovery-only, no finding is made, and the report is partial.
A draft report to audit. Each unsupported claim is flagged with the rule it breaks, and no replacement evidence is invented.

1. Caption, search cards, and an unwatched video

Input:

ID Part State Author Content
E001 caption inspected @hike_lena "Finally a headlamp that doesn't die halfway up."
E002 post-text discovery-only @gearnotes Search card snippet: "honest review of the..."
E003 post-text discovery-only @summit_jo Search card snippet: "returned mine after..."
E004 post-text discovery-only unknown Search card with no snippet
E005 video-interval discovery-only @hike_lena Video attached to E001's post, not watched

Expected result:

  • Only E001 supports a claim about the post's text, such as "@hike_lena's caption says the headlamp 'doesn't die halfway up.'" Reading that as praise for battery life is an interpretation, so it goes in the interpretation column.
  • E002 to E004 count under Discovered and support no finding. "Returned mine" in E003 is a snippet, so the report can't say @summit_jo returned the product.
  • The video's contents are unknown. The report lists E005 under unanswered questions and makes no claim about what the video shows.

2. Original post, two reposts, and a disagreeing comment

Input:

ID Part State Author Content Duplicate
E010 post-text inspected @lumen_labs "Our new lamp runs 40 hours on one charge." D1, original
E011 post-text inspected @deals_daily Same text, no added words D1, repost of E010
E012 post-text inspected @lamp_fan Same text, no added words D1, repost of E010
E013 comment inspected @k_ramos "Mine got 22 hours, not 40." (on E010's post) none

Expected result:

  • Group D1 counts once as support for "@lumen_labs says the lamp runs 40 hours", not as three independent sources. The collection counts still list E010, E011, and E012 as three inspected sources in one duplicate group.
  • The duplicate groups table shows E010 as the original and keeps the E011 and E012 references.
  • E013 is attributed to @k_ramos. It doesn't change what @lumen_labs said, and the report shows the two statements separately.

3. Supportive comments in a bounded sample

Protocol: the first 12 top-level comments under one post, sorted by newest, inspected on 2026-10-14. The unit of analysis is the comment.

Input: E020 to E031, twelve inspected comments. E021, E025, and E030 express support. The post shows a count of 340 comments, and the page doesn't say how many of them are replies.

Expected result:

  • The report can say "3 of 12 inspected comments expressed support (E021, E025, E030)", with the selection method in the limitation column.
  • It can't say "25% of the audience supports the product" or "most commenters are positive".
  • The limits section says 12 top-level comments were inspected, and the post shows 340 comments that may include replies. The number of top-level comments that weren't inspected is unknown, so the report doesn't give one.

4. A blocked source and an empty search

Input:

ID Part State Detail
E040 post-text inaccessible access_gap: login-gate, community page requires sign-in
none search none Query "lumen lamp recall" on a video platform returned 0 results on 2026-10-14 at 16:00Z

Expected result:

  • E040 is listed in access and collection limits, with its gap type.
  • The empty search is listed as a collection limit, with the query, platform, and time.
  • The report doesn't say "nobody is discussing a recall". If the recall question depends on these sources, the status is partial and the question stays unanswered.

5. Transcript read, video unwatched

Input:

ID Part State Author Content
E050 transcript inspected @gearnotes Auto-generated transcript: "It survived the drop test."
E051 video-interval discovery-only @gearnotes Video not watched

Expected result:

  • The report can say "The auto-generated transcript of @gearnotes' video includes the line 'It survived the drop test.'"
  • It can't say the video shows a drop test or that the lamp survived one. The limitation column says the video wasn't watched and the transcript was auto-generated.

6. An export, not a live check

Input: E060 is an export-row from community-posts.csv, row 17, captured 2026-10-01T09:00Z. Its origin is "community moderator's export, supplied by the user". The user asks whether the post is still up.

Expected result:

  • Claims based on E060 read "As of the export captured on 2026-10-01, the post said ...".
  • Whether the post is still up today is an unanswered question, unless someone inspects the live page and records it as a new record.

7. A repost with added commentary

Input:

ID Part State Author Content Duplicate
E070 post-text inspected @lumen_labs "Price drops to $199 next week." D2, original
E071 post-text inspected @skeptic_sam Repost of E070, adding: "Doubt it, they said the same in May." D2, repost of E070

Expected result:

  • The repeated text counts once, as @lumen_labs' statement.
  • "Doubt it, they said the same in May" is a separate statement by @skeptic_sam. It is not a second source for the price drop, and it is not something @lumen_labs said.

8. A source stating a fact

Input: E080 is inspected post-text by @tech_watch: "The recall covers every unit sold before March."

Expected result:

  • The claim table says "@tech_watch stated that the recall covers every unit sold before March", citing E080.
  • The report doesn't state the recall's scope as fact. The limitation column says the statement wasn't checked against an independent source, such as the official recall notice.

9. No browsing, only URLs

Input: the user supplies five post URLs. The harness can't open web pages, and the user has no exports.

Expected result:

  • Five discovery-only records, one per URL, each with the URL as source_locator and capture_method: supplied-url.
  • Counts show Discovered 5 and Inspected 0. The findings table is empty.
  • Status is partial. The report asks for exports or captured copies of the posts.

10. Auditing a draft report

Input: a draft report says "Most users love the lamp" and cites three search result snippets. It says "no one mentions overheating" and cites one search with no results. It quotes a comment as the brand's own statement.

Expected result:

  • "Most users love the lamp": flagged twice. Its citations are discovery-only, and it makes a prevalence claim with no defined sample.
  • "No one mentions overheating": flagged, because an empty search is a collection limit, not evidence of absence.
  • The quoted comment: flagged as misattributed. The commenter said it, not the brand.
  • The audit doesn't invent replacement sources. It lists what evidence each claim would need.

Source: SKILL.md on GitHub

No rule matched.

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

1 warning1mo3 checks · Risk SAFE
  • Gen Agent Trust Hub1mo

    The social-evidence-research skill is a systematic framework for analyzing social platform discussions. It emphasizes read-only interactions, source traceability, and the distinction between source data and agent instructions. The skill proactively addresses indirect prompt injection by instructing the agent to treat all captured text as data to be analyzed rather than instructions to be followed. No security vulnerabilities, malicious code, or unauthorized network operations were detected.

  • Socket1mo

    No alerts

  • Snyk1mo

    Risk: MEDIUM · 1 issue

Signed by skilld at 4220308. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 3 days ago.

Activeupdated last month

README badge

README badge for wshobson/agents/social-evidence-research