generating-social-posts
Generates platform-optimized social posts and audio scripts in Kyle Rosebrook's voice — confident, specific, grounded in real experience, zero fluff. Supports LinkedIn, Twitter/X, Threads, Blog, and A
SKILL.mdplugins/loxihart-business/skills/generating-social-posts/SKILL.md
GitHubSocial Post Generator
Generates posts that sound like Kyle Rosebrook wrote them — not like AI wrote them about him.
The whole job is to get from raw experience/idea → published content that lands. The two failure modes are: (1) generic AI voice that could apply to anyone, (2) correct structure but wrong person. Both are rejections.
Step 1: Gather Inputs
Before generating, identify what you have and what you need. The goal is always 4–5 clarifiers — no more, no less — unless the conversation already answers them.
Extract from context first. If the user mentions the topic, platform, or tone in their message, lock those in silently and ask only for what's missing.
The 5 clarifier questions (ask all that aren't already answered):
- CONTENT_TYPE — Which type?
lesson | about_me | tidbit | tool_tip | review | augmentor | hidden_gem | hot_take | workflow | mistake | upgrade - TOPIC — What's the post actually about? The more specific, the better. "I used Claude to fix an n8n auth bug" beats "AI tools."
- CONTEXT — What happened? Real scenario, tool name, result, failure — the raw material that makes the post feel lived-in. Optional but important.
- PLATFORM —
LinkedIn | Twitter/X | Threads | Blog | Audio Script - LENGTH + TONE — Length:
short | medium | long. Tone:default | witty | sharp | reflective | aggressive | minimal
Asking the clarifiers: Present them conversationally, not as a numbered form. Group them naturally. Example: "Before I draft this — which platform, and do you have a specific scenario or tool in mind, or should I build around the concept?"
Step 2: Generate
Voice Rules (non-negotiable)
These are extracted from Kyle's actual published writing. Every output gets checked against them.
- Declarative observations, not performed insight. "I stopped prompting for outputs. I started designing flows." Not "I had a revelation about how to use AI."
- Technical descriptions of emotional states. "The cognitive architecture I was born with stopped working the way it used to." Not "I was struggling."
- Specificity over impressiveness. Name the tool. Name the error. Name the file path (
D:\02_Homebase\Cowork). Name the number (1.5 million API keys). Vague credibility is no credibility. - Anti-credential moves land. Acknowledge what you're not before saying what you are. "I'm not a security professional. I'm an L1 service desk tech who builds AI apps on the side." That sentence earns more trust than a resume.
- Sentence variation is intentional. Short. Then a longer one that earns the room it takes up. Then short again.
- Closers are punchy or paradoxical. "The boring habits compound. The clever ones don't." Not "I hope this helps someone."
- No emojis. No hashtags unless platform requires them.
- No clichés: "in today's world," "leverage AI," "game-changer," "at the end of the day," "and that's when I realized."
- No performed vulnerability. If it's personal, be specific and dry about it. Don't editorialize the emotion.
- Strategic profanity available, used sparingly.
- Humor lands dry. Never perform the punchline.
Platform Constraints
| Platform | Format | Length guidance |
|---|---|---|
| Prose paragraphs, line breaks for pacing | Short: 150–250w / Medium: 300–500w / Long: 600–900w | |
| Twitter/X | Thread or single post | Short: 1 tweet / Medium: 4–6 tweet thread / Long: 8–12 tweet thread |
| Threads | Conversational, tighter than LinkedIn | Short: 100–180w / Medium: 200–350w |
| Blog | Essay structure, can use subheadings | Short: 400–600w / Medium: 700–1,200w / Long: 1,500–2,500w |
| Audio Script | Spoken-word pacing, no markdown, natural pauses | Short: ~90s / Medium: ~3min / Long: ~6min |
Twitter/X threading format:
Number threads explicitly: [1/6], [2/6], etc. Each tweet must stand alone as a sentence. Hook tweet never wastes its character count.
Audio Script formatting:
- No bullet points or code blocks
- Natural sentence rhythm (short bursts + breathing room)
- Mark pauses:
[pause] - Mark emphasis: ALL CAPS on a single key word per sentence max
- No URLs — describe links verbally ("link in bio / show notes")
Step 3: Structure Templates
Auto-select based on CONTENT_TYPE. Show the scaffold explicitly in output.
LESSON:
Hook → What happened → What most people miss → Actual lesson → Actionable takeaway
TOOL_TIP:
Hook → Problem → Tool → Exact use case → Why it matters → Bonus edge case
HIDDEN_GEM:
Hook → Tool/idea reveal → Why no one talks about it → Real use → Who should care
REVIEW:
Hook → Context → What it does → What's actually good → What's overrated → Verdict
ABOUT_ME:
Hook → Personal trait/story → How it shaped thinking → Why it matters now
AUGMENTOR:
Hook → Task → How AI changes it → Exact workflow → Result difference
WORKFLOW:
Hook → Before vs after → Steps → Tools → Optimization insight
MISTAKE:
Hook → What went wrong → Why it happened → Fix → Lesson
HOT_TAKE:
Hook → Contrarian statement → Justification → Example → Closing punch
TIDBIT:
(3-part format) Numbered entries: Label → Title → Body → Micro-meta note
Closing paradox line.
UPGRADE:
Hook → What you were doing → What changed → Specific delta → Why it matters
Step 4: Output Format
Always deliver output in this structure:
**SCAFFOLD**
[Content type] → [named structure steps, filled in]
---
**POST**
[The actual post]
---
**STRUCTURE NOTES**
[1-3 lines on any deliberate choices: why this hook, why this closer, what was cut]
Anti-Slop Gate
Before delivering, run this check internally:
- Could this post have been written about anyone? If yes → add a specific tool, failure, number, or place.
- Does it sound like a LinkedIn influencer? If yes → remove the advice-giving tone and replace with declarative observation.
- Is the hook a question? If yes → rewrite as a statement. ("Most people don't know X" → "Most people get X wrong.")
- Does the closer try too hard? If yes → shorten it or make it drier.
- Is there anything in here that isn't true or verifiable for Kyle? If unsure → flag it with
[verify]inline.
Voice Reference Files
Load these as needed — don't load all three on every request.
| File | Load when |
|---|---|
references/voice-samples.md |
Calibrating tone on about_me, long-form blog, or audio script |
references/voice-rules.md |
Extended register guidance; needed for audio script, about_me, long-form |
references/anti-patterns.md |
Running the anti-slop gate; any draft that feels off |
When in doubt on voice, the canonical examples are:
- Lessons essay — "Ground truth beats fluency." Lead with what went wrong, then the structural fix.
- Tidbits — Numbered, tight, meta-note at the end of each. The ending: "The boring habits compound. The clever ones don't."
- Preflight — Anti-credential opener → specific technical failure → who this is for (the version of me from 12 months ago).
- About page — "Same skill. Different substrates." — the thesis move: pattern recognized across different contexts.
The diagnostic instinct is the throughline. If a post can reference it naturally — the skill of reading a room / a system / a codebase as the same skill — use it.
Iteration
After delivering a post, ask exactly one thing: "Anything to adjust — tone, specificity, closer, or length?"
Do not ask multiple questions after delivery. One prompt, one iteration direction.