Last month I needed to put out three fresh posts a week for a niche site I run, and writing each one from scratch was eating up my mornings. I had a vague idea of using AI to help, but every tutorial I found either skipped the messy parts or assumed I already had a developer on staff. After a few failed attempts, I cobbled together a working pipeline that now spits out a first draft in under ten minutes, leaving me time to edit and publish.
Why does the AI keep sounding generic?
Most guides tell you to just feed a topic to ChatGPT and call it a day. The output reads like a Wikipedia summary written by a robot that has never met a human. The problem isn’t the model; it’s the prompt. If you ask for “a blog post about remote work,” you get a bland list of pros and cons that could belong to any article from 2020. The fix is to give the AI a clear role, a target audience, and a specific angle.
Here’s the prompt I use with Claude 3.5 (the first time I mention the tool I bold it):
You are a seasoned freelance copywriter who writes for indie hackers. Write a 1200‑word blog post titled “How to price your first SaaS product without scaring away early users”. The tone is conversational but data‑driven. Include two short case studies, one bullet list of pricing models, and end with a three‑question checklist for the reader. Use UK spelling.
That prompt costs roughly $0.004 per call on the Claude API (based on the current $8 per million token rate). For a 1200‑word post the input and output together stay under 800 tokens, so each generation is under a cent. I run three of these a week, which adds up to less than $0.15 in AI fees.
What most guides get wrong
They treat the AI as a finished Writer instead of a raw material supplier. The first draft is never publish‑ready; it needs shaping. Many tutorials skip the editing loop and tell you to copy‑paste the output straight to your CMS. That leads to bland posts, factual slips, and a tone that feels off‑brand. The real work begins after the AI spits out its text.
My process splits the job into three stages: outline generation, draft writing, and human polishing. Each stage gets its own dedicated prompt and its own temperature setting. Keeping the stages separate makes it easier to spot where the AI went off track.
How to debug when this breaks
When the output starts repeating phrases or veering off topic, the first thing I check is the prompt length. Claude 3.5 has a hard limit of 32k tokens, but if I accidentally concatenate previous outputs into the next prompt, the model sees noise and starts to echo. I keep a simple log file that stores the exact prompt sent each run; comparing a good run to a bad one usually reveals a stray line break or an extra space that threw off the token count.
If the tone feels wrong, I look at the temperature setting. For outline generation I use 0.2 to keep it deterministic. For the draft I raise it to 0.7 to add creativity. If I forget to reset the temperature between stages, the outline becomes too wild or the draft too stiff. A quick glance at the log fixes it.
Finally, I watch the token usage. If the output is truncated mid‑sentence, I know I exceeded the max output tokens. I either lower the max_tokens parameter or ask the model to continue in a second call. The latter works well because I can append “Continue from where you left off” and keep the same context.