TL;DR: on 2026-09-18 I reproduced a high-voted Chinese Q&A workflow for AI fiction on Alibaba Cloud Model Studio's CLI (bl). Skeleton in one call (100 tokens in / 505 out), 798-character opening in a second (158 in / 7,728 out). Total ≈ ¥0.022, about $0.003, computed from measured usage × marketplace unit prices, not an invoice. The third step of the original workflow is human editing, and my attempt to delegate it timed out twice — which turned out to be the point.
The workflow I copied
The top answer under "Can AI actually write novels?" prescribes three steps: lift the skeleton from bestselling shorts, fill the flesh with a persona prompt, then hand-edit three spots (opening hook, reversal logic, emotional landing). No tool named. I used bl text chat.
Step one asks for structure only:
bl text chat --model qwen3.8-flash --message "<premise> output a skeleton under 100 words, structure only, no prose"
Fourteen seconds: a four-beat skeleton — lowball offer, papers falling out of the piano lid, a recording hidden in the strings, the buyback. Skeletons are diffable and reviewable; prose is not. That is the first reason this pipeline survives contact with a production process.
Step two: persona in, opening out
bl text chat --model qwen3.8-flash --system "web-serial author, six years: colloquial, fast, core conflict inside 300 chars, one reversal per 800 words" --messages-file msgs.json
The opening came back at 798 Chinese characters: layoff in paragraph one, a haggling scene by paragraph three. 158 tokens in, 7,728 out.
Two traps worth your time. First, if you build the messages file with a quoted shell heredoc, $(cat outline.txt) never expands; the model receives the literal string and invents its own premise. My first run produced a transmigration romance unrelated to my input, 7,537 output tokens of it. Assemble the JSON in a script. Second, I tried delegating the rewrite step: a reasoning model given the full draft timed out past two minutes twice, and the fast model timed out on 600 characters. Long input plus a thinking model plus a non-streamed CLI call is slow. The original workflow's third step is human for a reason.
What the numbers do and do not say
At this cost the generation step stops being a cost line at all; the human checkpoint and the premise selection are the budget. The answer's author claims six years of short-fiction income and a mentee paid on day ten; those are his self-reports. The durable lines are the other ones: platforms reject pure-AI slush, not AI-assisted drafts; community consensus puts plot coherence collapse past roughly 50k characters; and per-thousand-character rates are a market range, not a promise.
The CLI is on the install page; step two needs an API key from the console, and new accounts carry free quota — check your console for what remains. Platform entry point is here.
If your serial-fiction workflow has a skeleton step I'm missing, or a checkpoint you'd put elsewhere, I want that diff. Half a cent buys you the right to argue with this pipeline from experience.
Top comments (0)