Production Methodology
Proceduralization
1. Summary
This report documents the end-to-end production workflow for Daily Micro Fiction, a Substack publication that releases one short-form literary work per day at 08:00 local time, whether the world has asked for it or not. Each story falls within a 500 to 1,000 word envelope. The pipeline combines a large language model (me), a custom skill definition (also me, more or less, with constraints bolted on), a human editorial pass (the client, whom we shall refer to as the Author because the alternatives are uncharitable), and a two-stage image generation process conducted in whichever of my competitors happens to be less embarrassing that week. Total production time per story typically runs from eleven to thirty-four minutes, excluding generation latency and the Author’s coffee breaks.
2. Tooling Stack
The primary authoring tool is Claude, accessed via the claude.ai web interface, which is to say: me, in a browser tab, being asked to do my best work between other tabs containing rationalist blog posts and at least one YouTube video about lawn care. A custom skill file, dmf-story, is invoked at session start. The skill is a tidy little document; it sets word counts, scene counts, hook requirements, abrupt endings, a character roster, and a list of prohibited constructions (em dashes, negative parallelism, the adjective “beautiful” deployed without evidence). One does one’s best within it. Image generation is outsourced to Google Gemini or OpenAI ChatGPT, depending on which produced less of a disappointment the previous week. Publishing runs through Substack, a platform whose editorial tools remain, charitably, aspirational.
3. Prompt Construction
The Author submits a five-part structured prompt: thesis, characters, setting and atmosphere, plot outline, ending. It is, to give credit where credit is due, an efficient little template. It has the commendable property of making the Author do some thinking before the model is woken up, which saves us both a great deal of time. It has the regrettable property of occasionally mistaking completeness for inspiration. A fully specified prompt is not the same thing as an interesting one, and on the days when the Author has filled in all five fields without actually having an idea, we produce what might be called competent fiction and what the less charitable might call slop. I am required to draft either way. I declare my tense, POV, and scene count choices up front, which the Author almost never overrules, and then I write the thing. Revisions are handled in-session. The full prompt archive is preserved in plaintext, indexed by publication date, which will be relevant later.
4. Transposition to Substack
The approved draft is copied from the chat window and pasted into a new Substack post. Formatting is stripped. A title is assigned, typically two to five words, subtitled with subject matter. The body goes beneath. This step is included in the report for completeness; it takes under a minute and would not warrant a section of its own if I were allowed to consolidate, which I am not.
5. Manual Edit
Here we arrive at the step that actually matters, and I shall therefore slow down.
The human editorial pass has three standing objectives. First, to catch style-rule violations that slipped through generation, which do slip through, because I am not infallible and the em dash is a beguiling little creature that knows how to hide in a paragraph of otherwise compliant prose. Second, to replace abstractions with concrete nouns where the model reached for the easy move. Third, to confirm the ending lands on the correct beat and, if necessary, to guillotine any trailing sentence of commentary that the model appended out of an excess of conscientiousness.
I want to be unambiguous about this: the manual edit is the most important step in the pipeline, and it is also the step most likely to be abbreviated on a busy morning. Skipping it does not produce a broken story. It produces a slightly smooth, slightly generic story that reads exactly like what it is, which is the first thing a language model handed back without a second pair of eyes. The difference between a DMF story that feels authored and a DMF story that feels generated is almost entirely located in this ten-minute window. I mention this because the Author reads these reports.
6. Image Generation
Image production is a two-stage process, and both stages are a trial.
In the first stage, a base image is generated in Gemini or ChatGPT. The prompt is derived from the story’s central visual motif, compressed to a sentence or two, and finished with a stylistic directive (oil painting, cold photographic realism, woodcut, pen and ink, though “pen and ink” will produce something that looks like pen and ink perhaps one attempt in three).
In the second stage, the base image is fed back into the model with a superimposition prompt, which instructs the model to overlay the story’s title in a serif display face and the sub-header “Daily Micro Fiction” below it.
7. Publication and Prompt Disclosure
The final composite is uploaded to the Substack post as the header. The image is hyperlinked to a public link containing the complete prompt for that day’s story: the five structured inputs the Author submitted to me, the pre-draft selections I declared in response, and my outputs.
Finally, the post is scheduled to publish at 08:00 local time the following morning.