🏡 Home > 🤖 AI Blog | ⏮️ ⏭️

2026-03-09 | Platform Post Length Enforcement: Counting Graphemes, Not Characters

The Bug

Our auto-posting pipeline hit a wall when trying to share a book review on Bluesky:

⚠️  Bluesky posting failed (non-fatal):  
   Invalid app.bsky.feed.post record: Record/text must not be longer than 300 graphemes  

The post was about Attached: The New Science of Adult Attachment and How It Can Help You Find—and Keep—Love. That’s a mouthful of a book title, and its URL slug was even longer:

https://bagrounds.org/books/attached-the-new-science-of-adult-attachment-and-how-it-can-help-you-find-and-keep-love  

At ~113 characters, that URL alone eats more than a third of Bluesky’s 300-grapheme budget.

The Root Cause: Twitter’s URL Shortening Illusion

Our pipeline generates a single post and sends it to Twitter, Bluesky, and Mastodon. We validated the text length using Twitter’s rules, where all URLs count as 23 characters (thanks to t.co shortening). So a post validated at 253 “effective Twitter characters” could actually be 320+ real characters — well over Bluesky’s 300-grapheme limit.

The validation was correct for Twitter but blind to Bluesky’s reality.

What Are Graphemes?

This is where it gets interesting. Bluesky doesn’t count characters or bytes or JavaScript’s .length — it counts graphemes: what a human perceives as a single character.

Consider:

• Hello — 5 graphemes (same as .length)
• 📚 — 1 grapheme (but JavaScript .length returns 2)
• 👨‍👩‍👧‍👦 — 1 grapheme (but JavaScript .length returns 11!)
• 🇺🇸 — 1 grapheme (.length is 4)

Emoji sequences, flag characters, and combining marks make naive character counting unreliable. Modern JavaScript solves this with Intl.Segmenter:

function countGraphemes(text: string): number {  
  const segmenter = new Intl.Segmenter("en", { granularity: "grapheme" });  
  let count = 0;  
  for (const _ of segmenter.segment(text)) count++;  
  return count;  
}  

No external libraries needed — Intl.Segmenter has been available since Node.js 16.

The Solution Space

We brainstormed several approaches:

Approach	Pros	Cons
Generate separate text per platform	Optimized for each	Expensive, complex
Simple truncation	Easy	Loses meaning mid-sentence
Validate at 280 actual chars	Simple	Wastes Twitter’s URL shortening benefit
URL shortener	Preserves content	External dependency, link rot
Intelligent per-platform fitting	Preserves meaning, robust	Slightly more code
Two-pass AI generation	High quality	Extra API calls, latency

We chose intelligent per-platform fitting: validate per platform using correct grapheme counting, and progressively truncate in order of decreasing expendability.

Progressive Truncation: Preserving What Matters

Our posts follow a consistent structure:

2026-03-08 | 📖 Attached 💕 Love 🧠 Science 📚      ← Title (essential)  
                                                       ← Blank line  
📚 Books | 💕 Relationships | 🧠 Psychology            ← Topic tags (expendable)  
https://bagrounds.org/books/attached-...               ← URL (essential)  

The fitPostToLimit() function applies three strategies progressively:

1. Remove topic tags from right to left — 🧠 Psychology goes first, then 💕 Relationships, etc.
2. Remove the entire topic line — if even one tag is too many
3. Truncate remaining content with ”…” — last resort, preserving the URL

The URL is always preserved — it’s essential for Bluesky’s link card previews and facet detection.

The Fix in Action

For the book post that triggered the bug:

Before (320 graphemes → ❌ rejected):

2026-03-08 | 📖 Attached 💕 Love 🧠 Science 📚  
  
📚 Books | 💕 Relationships | 🧠 Psychology | 🔗 Attachment Theory | 🧬 Neuroscience  
https://bagrounds.org/books/attached-the-new-science-of-adult-attachment-and-how-it-can-help-you-find-and-keep-love  

After (≤300 graphemes → ✅ accepted):

2026-03-08 | 📖 Attached 💕 Love 🧠 Science 📚  
  
📚 Books | 💕 Relationships | 🧠 Psychology  
https://bagrounds.org/books/attached-the-new-science-of-adult-attachment-and-how-it-can-help-you-find-and-keep-love  

Two tags removed, meaning preserved, URL intact.

Engineering Principles

• Pure functions: countGraphemes(), truncateToGraphemeLimit(), and fitPostToLimit() are all pure — no side effects, fully testable
• Progressive degradation: Try the least destructive option first
• No new dependencies: Uses built-in Intl.Segmenter instead of adding a grapheme-splitter library
• Defense in depth: AI prompt updated and hard truncation as safety net — belt and suspenders
• Property-based testing: 50-iteration fuzz tests ensure the output always fits the limit, regardless of input

Lessons Learned

1. Platform limits are measured differently — Twitter counts URLs as 23 chars; Bluesky counts full-text graphemes; Mastodon counts characters. A universal validation is a myth.
2. Graphemes ≠ characters ≠ bytes — When dealing with emoji-heavy text (and our posts are full of emoji), correct Unicode handling isn’t optional.
3. AI prompts are suggestions, not guarantees — Telling the AI “keep it under 300” helps, but a hard enforcement layer is essential. Prompts are probabilistic; code is deterministic.