Analyze

Skill gokul-kulkarni/content-studio-plugin/skills/analyze

Analytics & insight phase — turn Substack, Medium, and LinkedIn stats into a data-backed audit of content quality AND outreach, with concrete next actions mapped to the writing/distribution skills. Ingests normalized records from the substack-stats / medium-stats / linkedin-stats skills (or manual data), finds what's working and what's leaking reach, reads the cross-platform funnel, and ends with a prioritized action list plus a 'what to write next'. Use whenever someone asks how to improve their outreach, why a post under- or over-performed, what to write next, or wants an analysis/audit of their publishing across platforms. This is Phase 5 — it closes the loop back into explore/write/distribute.From its SKILL.md

Install

npx -y skills add gokul-kulkarni/content-studio-plugin --skill analyze

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

7.0 KB, as published. Nobody here has run it

Analyze & Improve

Turn the numbers for $ARGUMENTS into decisions. This skill owns Phase 5 (Measure) of the pipeline: it reads what happened across LinkedIn, Substack, and Medium and tells the person what to do differently — then routes each fix back to the skill that executes it.

Step 1 — get the data

Pull stats first via the gathering skills — substack-stats, medium-stats, linkedin-stats — each of which parses the analytics exports the person downloads from the platform (Substack ZIP / Dashboard Stats, Medium GraphQL / Stats, LinkedIn XLSX), with an optional public fetch for inventory. If the person already has exports/pastes, ingest those directly. Aim for at least two platforms before drawing cross-platform conclusions; one platform is fine for a single-platform audit.

The normalized record (canonical schema — the gathering skills produce this)

Per post (PostRecord), nulls where a platform lacks a field:

field	meaning
`platform`	substack / medium / linkedin
`id`, `title`, `url`, `publishDate`	identity + timing (ISO date)
`format`	newsletter / article / listicle / carousel / post
`topics`	tags / subject
`impressions`	times shown (LinkedIn)
`views`	views / opens (Substack, Medium)
`reads`, `readRatio`	full reads and reads÷views (Medium's quality signal)
`reactions`	likes / claps / reactions, one number
`comments`, `shares`	comments; shares / reposts
`clickThrough`	outbound link clicks, where reported
`audience`	free / paid / everyone
`source`	export / public (confidence + coverage flag)

Plus one AccountSummary per platform: subscribers, paidSubscribers, followers, subscriberDelta, followerDelta, openRate, rangeDays, source.

Normalize before comparing. Raw counts favor whoever has the bigger audience. Compute rates — engagement ÷ reach (reactions/impressions or reactions/views), read-ratio, click-through ÷ reach — so a small-audience post that punched above its weight stays visible. Flag any metric that's null or source: public so a coverage gap isn't read as a zero.

Watch these measurement traps (each drawn from real data):

Reach ≠ right reach. The single highest-impressions post is often a personal/off-topic one that spread through a legacy network at a low engagement rate — big number, wrong audience. Rank by engagement rate; call out such outliers instead of letting them define "what works."
Substack open rate is inflated by Gmail's image proxy (opens logged as US regardless of location). Treat it as a directional title signal; for real engagement use unique-subscriber opens from the export.
Subscriber/follower growth is usually publish-day driven. Map each signup's date to the post that ran that day (Substack email_list.csv created_at, LinkedIn daily new-followers). If every publish converts, the growth lever is volume of top-of-funnel, not a better post — say so.
Demographics are the ICP. LinkedIn's demographics sheet (location, seniority, title, industry) tells you who you're actually reaching — check the content matches them, and name the ICP explicitly.

Step 2 — the audit (do both halves)

A. Content quality — what wins and why

Top vs bottom quartile by normalized engagement. What do the winners share — topic, format, angle, title/hook shape? What do the losers share?
Titles & hooks: which openings drove reach (impressions/views) vs which drove depth (read-ratio, comments). A high-view / low-read-ratio piece is a title that overpromised — a writing fix, not a reach one.
Format: carousel vs essay vs listicle vs HTML article — which format serves which topic best.
Series lift: do posts in a series outperform standalones? Is the callback/tease actually compounding?

B. Outreach / reach — where growth comes from and where it leaks

Timing & cadence: publish-time vs performance; is there a consistent, repeatable schedule? Gaps and bursts both hurt the recommendation engines.
Reach leaks: high impressions + low click-through, or high views + low read-ratio, or high reach + low follower/subscriber conversion — name the specific leak, not just "engagement is low".
Conversion: subscriber/follower delta against the posts that ran in the window — which pieces actually grew the audience.

C. Cross-platform funnel

Read the three platforms as one funnel: LinkedIn (discovery) → Substack (subscribe) → Medium (depth). Which topics win on which platform? Is the same thesis carrying across, or drifting?

Compute the actual conversion rate, not a vibe: LinkedIn members-reached → Substack subscribers. A large reach with a tiny subscriber delta is the leak, and it's usually two fixable things — no subscribe CTA / link in the LinkedIn post (links belong in the first comment), and no link out of the Substack email.
Discovery is not evenly distributed. One platform typically carries almost all reach (often LinkedIn); another (often Medium) may get near-zero organic discovery despite a decent read-ratio — meaning its content is fine but nothing feeds it. Don't read low Medium reach as low Medium quality; the fix is driving traffic to it (canonical + links from LinkedIn/Substack), or consciously de-prioritizing it.

Step 3 — recommendations, mapped to the skills that execute them

Every finding must become an action pointed at a specific skill:

Framing / topic selection / angle → explore (re-lock a stronger spine or pick better topics)
Hook / title / format / structure → write-linkedin · write-substack · write-medium
Timing / tags / cross-posting / canonical / launch Note → distribute
Restructure a specific underperformer → refine

End with:

A prioritized action list — highest-leverage first, each tied to a skill and a why-from-the-data.
"What to write next" — 2–3 concrete topic/format/platform bets grounded in what the winners share.

Optional — trend over time

Offer to save the normalized dataset (e.g. ./content-studio-stats.json in the project, or the scratchpad) so the next run can show trend-over-time (are the recommendations working?). Opt-in only; never auto-upload. Keep all analytics local — they're personal data.