If you've ever asked ChatGPT for "10 tweets about productivity," you've seen the problem: the output is technically correct, vaguely useful, and obviously not yours. It reads like AI because it was written by AI with no context about you.
Growly's suggestion engine fixes this by tuning every prompt with three things the generic interface never sees: your past posts, your declared style, and your performance signal.
The three inputs
1. Past posts (your voice)
We feed the model a sample of your recent Threads posts — verbatim. Length, capitalization, punctuation habits, sentence rhythm, how often you ask questions, whether you use em-dashes or hyphens. The model is asked to produce text that matches that surface texture, not generic "social media post" texture.
This is the biggest single lever. A model that's seen 30 of your posts can produce a 31st that reads like yours. A model that's seen zero will produce slop.
2. Declared style hints
You can tell the engine things it can't easily infer:
- What topics you cover, with priorities.
- What you don't talk about (politics, specific competitors, etc.).
- Tone (irreverent, serious, technical, casual).
- Formality (lowercase-everything vs proper sentence case).
These ride on every prompt as constraints.
3. Performance signal
This is where Growly differs from a pure styling tool. We look at which of your past posts actually performed — by views, replies, likes — and bias the suggestions toward post-types that worked for you historically. If poll-style posts dominate your top-10 by reply count, the engine leans poll-flavored. If short single-line takes dominate by view count, the engine biases short.
What we don't send
A few intentional omissions:
- No DMs, no drafts you didn't publish. Only public-Threads content is in the context.
- No personally identifying info beyond your Threads handle. No email, no IP, no payment info.
- No real-time data. We don't query Threads' API at prompt time; we use what we've already stored.
Why generic ChatGPT prompts fail
If you copy a prompt like "give me 10 Threads posts about marketing for solo founders," ChatGPT produces output averaged across the internet's voice. The result reads like every other LinkedIn motivation post because that's what the model has seen most of.
To get usable output from a raw chat interface, you'd need to:
- Paste 30 of your past posts as context.
- Describe your style explicitly.
- List your taboo topics.
- Specify the platform's character limit.
- Ask for variety.
- Reject the first 7 outputs.
That's a 15-minute prompt-engineering exercise per session. Doing it inline as a feature is what Growly automates.
What it's not
A few honest limits:
- Not a replacement for original thought. The engine is a brainstorming partner, not an idea factory. The best output comes when you bring half an idea and let it complete the framing.
- Not infallible. Suggestions can miss the brief. The cheap iteration cost (one click for a new batch) is the answer, not perfection per-output.
- Not training on your data. We send your posts to the model per-request; we don't fine-tune a personal model. The model itself is stateless across users.
How to use it well
The pattern we see in heavy users:
- Open the composer with a vague intent ("write something about today's release").
- Ask the AI for 5 angles.
- Pick the one that resonates; either ship as-is or use it as a draft to rewrite in your own typing.
The output is the seed, not the final tree.
What's coming
Iterating in two directions: deeper style modeling (the engine learns from corrections, not just past posts) and topic-cluster targeting (suggestions tuned to the specific slot's tag rather than your global style). Both ship later this year.
If you've been burned by generic AI prompts, give a style-tuned one a try. The output curve is materially different.