Your Best Posts Are the Training Data: Distilling a Claude Skill From Your Own Hits

You wrote four hundred posts over three years. Twelve of them did numbers — real numbers — and you have a dim sense of why, but you couldn't write down the actual formula if someone paid you. Meanwhile every new post starts from the same blank-page ritual: think of topic, remember what worked, half-recall a hook shape, draft, hope.

The uncomfortable truth: the data answering "what works for me" already exists, sitting in your own post history with engagement metrics attached. And the boring-but-effective move is to have AI mine it — feed your best-performing posts to Claude, have it extract the patterns you can't see, and hardcode the result into a reusable skill. Thirty to sixty minutes of setup, and every future post starts from your proven formula instead of a blank page.

This is the pipeline: raw history → strategy report → reusable skill. (The companion piece on this blog covers the broader multi-role content pipeline; this one is specifically about training on your own successes.)

Step 1: Extract Your History Safely

With hundreds of posts, copy-pasting into a chat is both inefficient and error-prone. Use a data-extraction tool to pull your public posts into a spreadsheet (CSV/XLSX). One safety rule matters more than the tool choice: use a third-party scraper, never an automated bot on your own account. Bot automation on your profile risks platform penalties; third-party extraction of public data keeps your account safe and inside standard data-collection practice.

The dataset you want has three columns minimum:

One export, one file — clean and structured. This is the "training material," and its quality caps everything downstream.

Step 2: Generate the Strategy Report

Upload the spreadsheet — top performers only — and skip the vague prompt. "What's good about these?" produces a book report. A technical prompt produces analysis:

You are analyzing my best-performing LinkedIn posts
(sorted by engagement, top decile). Produce:

1. WINNING FORMATS — which structures dominate:
   story-led, list, contrarian take, how-to, data point?
   Count and rank them.
2. HOOK PATTERNS — the first 1-2 lines of the top posts:
   quote the actual openings, group them into hook types,
   note length and rhythm.
3. OUTLIER AUTOPSIES — the 3 posts that beat average by
   3x or more: what did each do differently? Be specific,
   quote the evidence.
4. VOICE MARKERS — sentence lengths, vocabulary habits,
   punctuation patterns that appear consistently.
5. FAILING COUSINS — patterns in my average posts that
   the top posts avoid.
Output as a structured report. Quote real examples for
every claim — no generic advice.

The two clauses doing the heavy lifting: quote the evidence (forces analysis, not invention) and outlier autopsies (the 3x posts carry more signal than the top-20 average). AI is genuinely excellent at this kind of pattern recognition across a dataset — finding the consistent structures that a human "sort of knows" but never articulates.

Step 3: Bake It Into a Reusable Skill

The report is analysis; the skill is the asset. Take the findings and convert them into a saved instruction set — a skill in Claude's terms, a persistent custom instruction anywhere else:

### /post — my content formula (from data analysis of top posts)

Structure: [winning format #1 from report], fallback [format #2]
Hook: open with [dominant hook type], under 15 words, no greeting
Body: 1 core idea, short paragraphs, one concrete example with numbers
Voice: [voice markers — e.g., direct sentences, contractions,
       occasional one-word paragraph for emphasis]
Banned: [patterns from "failing cousins" — e.g., "In today's
        world" openers, list posts without a takeaway]
Process: give me 3 variations on [TOPIC], each using a different
winning hook type. Flag any variation that drifts from formula.

From then on: "Write me a post about [topic] using the format we established" — and the output arrives inside your proven formula instead of the generic-professional average. You skip the re-explaining, every single time.

Manual Drafting vs the Trained Pipeline

The Three Traps

Garbage in, garbage out. Feed the model two hundred average posts and it will learn to write average content with great confidence. Curate ruthlessly — top decile only. The dataset is the strategy.

The recycled-hook trap. A hook that worked for one topic doesn't transplant to every topic. The skill encodes hook types, not one magic opener stamped onto everything. Relevance to the audience still gates each draft — that judgment stays yours.

Over-reliance. The model replicates patterns; it doesn't have your stake in the ideas. The final edit — does this actually say something I believe, to people I know? — is the human layer, permanently. That's not a limitation of the technique; it's the part that makes it worth reading.

The Local Variant

The whole pipeline is text-and-tables, which means it runs locally too: the export and analysis work with a local model (Ollama) and a private spreadsheet when the post history touches client accounts or unreleased positioning. Nothing about your engagement data or your formula needs to reach a third-party server. Cloud for convenience, local for anything sensitive — the same split this blog applies everywhere.

Frequently Asked Questions

Does this work for platforms other than LinkedIn?

Yes — the logic is platform-agnostic. Anywhere you have a high volume of past content with engagement signals (X, Instagram, a newsletter), the same export → analyze → skill pipeline applies. Only the format taxonomy changes.

What does it cost?

A scraping tool subscription (free tiers exist) and AI usage for the analysis session. The one-time cost buys a permanent asset: the skill eliminates the re-explaining and speeds up every future post, so the setup pays back quickly at any regular posting cadence.

Do I need to code?

No. The pipeline is: run an extraction tool, upload a spreadsheet, paste a detailed prompt, save the result as an instruction set. Technical judgment helps write the analysis prompt, but no programming is involved.

How often should I redo the analysis?

Quarterly is plenty. Re-export, re-run the report, and diff it against the old skill — platforms shift what they reward, and your voice evolves. A skill built from two-year-old data slowly becomes a museum piece.

Wrap-Up

Your best content already told you what your best content looks like — it's sitting in the metrics. Export the history, feed the winners to Claude with a forensic prompt, and convert the report into a skill you own. You stop being a writer who rediscovers their formula every morning and become a strategist deploying one. Setup costs an hour; the formula pays every post after that.