nickvasilescu/nicks-stack/evidence-led-social-posts

Evidence-led social posts

Create social posts from current evidence rather than generic copywriting formulas. The goal is a sharp, credible post that fits the user's established voice and gives the attached artifact a clear reason to exist.

Trigger conditions

Use this skill when the user asks for:

• X/Twitter post or thread ideas
• launch, comparison, ranking, or announcement copy
• copy modeled on a creator's style
• outlier research before drafting
• a post promoting a video, benchmark, experiment, or product

For X research and account access, load xurl first. For the final anti-slop pass, load humanizer when available.

Workflow

1. Establish the factual spine

Extract only claims supported by the user's results or current source material:

• what was tested
• number of subjects or products
• whether conditions were comparable
• concrete tasks or use cases
• winner, runner-up, and surprise
• strongest caveat or close call
• attached asset and desired action

Do not invent test rigor, prices, speed measurements, dates, access conditions, or rankings. If the user supplies a ranking, preserve it unless live evidence contradicts it.

2. Research the original sources

When the user gives an account, creator, post URL, or topic, inspect that source directly. Do not substitute session history for the current X timeline.

Collect three evidence sets:

1. User voice: recent original posts, especially posts on the same topic.
2. Reference creator: recent original posts from the named creator.
3. Topic outliers: relevant posts from the current news window.

Exclude replies and reposts where possible. Capture post text, date, impressions, likes, replies, reposts, quotes, and bookmarks.

3. Identify outliers correctly

Do not rank posts by likes alone. For a quick within-account comparison, use a weighted engagement signal:

likes + replies + 2*reposts + 2*quotes + 3*bookmarks

Compute an outlier multiple against that account or query's median. Treat this as a heuristic, not a universal quality score. Also inspect impressions and bookmark rate because useful long-form posts often earn saves rather than replies.

Never compare raw engagement between a huge creator and a smaller user without normalizing. Learn structural patterns from the larger creator, not expected reach.

4. Extract structure, not imitation

From a reference creator, identify reusable mechanics:

• first-line hook type
• proof or authority placement
• specificity and numbers
• pacing and line breaks
• reveal timing
• use of lists
• ending and call to action
• relationship between post copy and attached media

Do not copy signature phrasing, capitalization habits, or entire sentence structures. State that the draft uses structural patterns rather than impersonating the creator.

5. Choose one dominant narrative

A strong post needs one central story. Useful patterns include:

• Reversal: “I thought X won. Then Y arrived.”
• Controlled experiment: “I gave N subjects the same task.”
• Unexpected podium: expected winner, close second, surprise third.
• Contrarian observation: the common evaluation method misses what matters.
• Practical verdict: best overall, fastest, best value, or best for a defined job.

Do not turn a post into release-note soup. Put the full lineup, methodology, and caveats in a first reply or thread when they weaken the opening.

6. Separate the hooks across title, post, and media

If a video title or thumbnail already carries the winner or reversal, the post should add proof, personality, or mechanism rather than repeat the same sentence.

A useful division is:

• post: narrative or personal reversal
• thumbnail/image: visual proof or experiment scale
• video title/link: explicit payoff
• first reply: full methodology, lineup, and discussion question

7. Draft distinct options

Produce 3 to 5 genuinely different directions, not cosmetic rewrites:

• recommended narrative version
• shorter version
• contrarian version
• creator-inspired structural version
• reply-oriented version when appropriate

Lead with the recommended draft and explain the choice briefly. Keep X posts within the user's actual account limits when known. Use a character counter rather than guessing.

8. Humanize and verify

Before delivery:

• read the copy aloud
• remove generic AI claims such as “changes everything” or “the landscape shifted”
• remove unnecessary hashtags and company-tag piles
• preserve natural contractions and imperfect rhythm
• check every proper noun and model spelling
• verify rankings and numbers
• confirm that superlatives are scoped, for example “best overall in my Hermes test” rather than “best model in the world”
• do not use em dashes for users who prohibit them

Recommended output format

1. Best post ready to paste
2. Why this angle in 2 to 4 bullets
3. Alternative versions with clearly different hooks
4. Suggested first reply for methodology, full lineup, or a question
5. Research signal listing the few outlier patterns that materially affected the copy
6. Confidence label

Do not bury the paste-ready copy beneath a long research report.

Pitfalls

• Generic “latest models this week” framing has no conflict or payoff.
• Listing every product in the opener creates metadata soup.
• A ranking without test mechanism feels arbitrary.
• Saying “destroyed” when results were close harms credibility.
• Copying a creator's voice too literally becomes impersonation.
• Raw likes are not an outlier analysis.
• Search results may contain engagement bait or fabricated claims. Use them as style signals only unless independently verified.
• Do not post, reply, quote, or upload media without the user's explicit approval.

Supporting references

• See references/ai-model-comparison-posts.md for the reusable research findings and post structures from frontier-model comparison launches.