PostSpark

The Insight That Started It All

Two founders, no code written, $1M in revenue in weeks. The “Stanley” story was the clearest validation of a simple thesis: the value in AI products is the prompt design, not the engineering.

Stanley’s insight: generic AI gives generic LinkedIn advice. What professionals needed was an AI that understood their specific story — their failures, wins, and contrarian beliefs — and generated content that could only come from them.

But after building v1, a harder question emerged: “Why not just use ChatGPT directly with the same prompt?”

The answer forced a real product decision: the competitive moat isn’t the prompt. It’s the accumulation.

v1 → v2: From Generator to System

v1 problem: Completely stateless. Brilliant output, but you re-entered your entire life story every session. Zero memory. ChatGPT with a prettier UI.

v2 thesis: The real value isn’t generating content — it’s remembering who you are and getting better at predicting what works for you.

What PostSpark v2 Does

Profile persistence — Your professional DNA is saved to localStorage the moment you generate. Next session: one click to continue, not a blank form.

Style learning — Every post card has a ❤️ like button. PostSpark tracks which formats and angles you respond to, then injects those preferences into the next generation prompt automatically.

2-week content calendar — Results aren’t just a list of 10 cards anymore. They’re distributed across Tuesday/Thursday/Friday slots for two weeks — the optimal LinkedIn posting rhythm for B2B audiences.

Profile version history — Last 3 profile snapshots saved. Click to restore any earlier version — useful when iterating on positioning.

Export / Import — JSON profile portability. Move between devices without a backend.

Profile completeness indicator — Live progress bar (0→100%) with color feedback, nudging users to fill the fields that most improve output quality.

The Core UX Flow

1. First visit — Fill your professional profile (name, title, industry, achievements, failure, background, hot take)
2. Return visit — Welcome back banner: “Use saved profile” in one click
3. Format selection — 8 content types; learned preferences shown as ”🧠 Learned from your likes”
4. Generation — 15-20s; thinking steps now include “Applying your style preferences”
5. Results — 10 post cards + calendar view; like what resonates
6. Next session — Profile pre-filled, preferences already learned

Why Prompts Alone Aren’t a Moat

The style learning system works by:

1. Storing liked posts as { type, hook } objects in postspark_liked
2. On next generation, getLikedPatterns() aggregates format frequency
3. The prompt gets a new block: "User prefers: Personal Story (3x), Hot Take (2x)"
4. The AI adjusts format distribution accordingly

A competitor can copy the prompt. They can’t copy your accumulated preference data.

Technical Design

Single HTML file, zero backend, zero server cost. Five localStorage keys manage all state:

• ps_apiKey — Fireworks AI key
• postspark_profile — Current profile JSON (9 fields + formats + savedAt)
• postspark_liked — Array of liked post objects
• postspark_versions — Last 3 profile snapshots

The Fireworks AI call goes directly browser → API endpoint. JSON response parsed with regex \[[\s\S]*\] + try/catch for resilience against malformed outputs.

UI has 5 states: hero → form (with welcome banner logic) → loading → results (with calendar) → error.

The Bigger Lesson

The real insight wasn’t “build a LinkedIn tool.” It was: the accumulation of personal data is the product. Prompts are replicable. What a user has built up over 20 sessions — their liked formats, their profile iterations, their content history — is not.

This is the pattern that scales: start BYOK and stateless, prove value, then add persistence as the moat deepens with every use.