2× publishing velocity. Voice intact.

The challenge

Pull My Frank is a personal editorial platform for an independent journalist writing across five categories — long dispatches, photo essays, short commentary, home reportage, and notebook fragments. The author publishes 10–15 pieces a month without an editor, copy desk, or assistant.

The bottleneck wasn't ideas — it was momentum. A piece started in a strong voice would stall mid-draft, and restarting it the next day meant rereading to re-find the register before a single new word could be written. Excerpt writing was a separate problem: producing sharp, two-sentence homepage teasers consistently took 20 minutes or more per piece, long after the piece itself was done.

The biggest concern with using AI was voice. The author's prose is distinctive — Southern literary, Borges-influenced, unhurried, grounded in specific place. Generic AI writing assistants produce prose that sounds like no one in particular. The requirement was an LLM that could continue a piece in the author's register without the author having to correct it.

The solution

Gilligan Tech Inc. integrated two lightweight AI features directly into the Pull My Frank CMS admin — no separate tools, no copy-paste workflow. Both features are surfaced as single-click buttons in the post editor the author already uses every day.

The first feature is an AI Assist button that appends a prose continuation to whatever the author has written so far. The continuation is generated by Qwen 3 running on the Gilligan Tech Inc. AI Platform, primed with a detailed voice profile that captures the author's register, sentence rhythm, thematic concerns, and sense of place. The author reviews the output, keeps what fits, and discards or edits what doesn't — but in practice, most continuations land cleanly enough to use as a draft floor.

The second feature is a Generate Excerpt button that reads the finished body text and produces a 1–2 sentence homepage teaser calibrated to the author's editorial style. The excerpt is written into the excerpt field and can be edited before saving. What previously took 20+ minutes of staring at a finished piece now takes seconds and one review pass.

Implementation

1. Voice profile engineering: Gilligan Tech Inc. developed a system prompt capturing the author's literary influences, characteristic sentence structures, thematic anchors (North Carolina, place, the overlooked ordinary), and stylistic red lines (no abstraction, no hedging, no machine-generic phrasing). The profile was iteratively tested against existing published pieces.
2. LiteLLM integration: A PHP endpoint (api/ai.php) was added to the CMS, routing requests to Qwen 3 via the Gilligan Tech Inc. LiteLLM gateway. The endpoint validates session auth and CSRF tokens before forwarding to the AI layer. Qwen 3 8B was selected for its fast inference and strong prose quality at this scale.
3. Editor UI hooks: Two AI action bars were added to the post editor — one below the body field, one below the excerpt field. Each is a single button with a spinner and status indicator. No framework dependencies; plain fetch() in vanilla JS.
4. TinyMCE continuity: The body editor uses TinyMCE. The AI Assist integration reads TinyMCE's content directly (via the editor API), appends the returned continuation as a new paragraph, and returns focus to the editor — maintaining formatting and cursor behaviour throughout.
5. Attribution: The site footer carries an "AI writing assistance by Gilligan Tech Inc." credit linked to the platform page, providing visibility for the AI layer to readers.

Results

Technology