Keep your voice in AI-assisted writing: two profile files, a three-pass cleanup that cuts 30 to 50 percent, then rebuild the voice by hand
A former freelance writer describes a three-step process for writing client and personal content with AI: set up values and voice files once, strip invented and generic text in three passes, then rebuild it with his own stance and stories.
Evidence: The author reports this. We have not checked it beyond reading the source.
The business problem
Clean AI output reads as lifeless, and models invent facts, stories and statistics that damage trust.
What was tried
Once, spend about 20 minutes on an interview that produces two files, one with values, origin story and firm positions and one with voice traits such as sentence rhythm and words he never uses. Then run a cleanup of three passes that removes invented content and list-style structure, cuts unnecessary sentences, and strips hedges, adverbs and AI tells, usually cutting 30 to 50 percent of the words, with any invented statistic flagged for checking. Rebuild the voice by hand in about 15 minutes, adding values, stance, stories and quirks, with the model asking or leaving labelled gaps instead of inventing. Keep an override log of the suggestions you reject.
What was reported (mixed)
He says a piece that took about eight hours takes about two. He reports his own lead quiz drew about 50 takers over six months with no sales calls, while a coach's similar assessment drew just over 60 and brought in more than $20,000 including coaching. Flat AI text scored 0 percent on a detector and the personal-voice version 7 percent.
Limitations
An interview in a curated series, with self-reported figures and the tools unnamed. He sells AI building services and promotes his own prompt library and tools. He says clean output can read as lifeless, oversight weakens over time, AI detectors are unreliable as a measure of quality, and fluent AI made his own weak quiz look finished so its quality was hard to judge. Relationship-based work such as his 23-founder survey cannot be automated.
What you need
An AI assistant, a willingness to write the profile files honestly, and time to review each output.
Sources
- AI Workflow Stories on Substack (interview with James Presbitero Jr.) ↗ Firsthand write-up, published September 15, 2026
Source published: September 15, 2026. Last reviewed here: October 13, 2026. Spot a mistake? Tell us.
Related workflows
Rank an AI-assisted article for a competitive keyword: study page one, rewrite 80 percent of the draft, then have Claude check for stuffing
A writer explains how an article reached first place on Google for a competitive phrase: analysing page-one results, drafting with a long prompt, rewriting most of it by hand, and using Claude to catch keyword stuffing and errors.
Write articles in your own voice with Claude: dictate, outline, score with three custom skills, edit heavily, then repurpose
An online marketer describes her weekly article workflow: dictating thoughts into a Claude project, outlining, scoring the draft with three custom skills, heavily editing in her own CMS, and repurposing each piece into social posts.
Use AI for the edges of a weekly newsletter, not the body: transcription, proofreading, subject lines and reader-reply themes
A newsletter operator records a voice note, writes the issue by hand, and uses AI only for transcription, a light proofread, subject line options and a monthly analysis of reader replies.
Set up Hermes Agent on a cheap VPS with Telegram, morning cron briefings and a reusable content research skill
A 14-day test of Hermes Agent on a small VPS, reached through Telegram, with memory files, a daily and weekly briefing, and custom skills for content research.