Memoir writer

The failure mode this is built against

Ask a language model to write a chapter of someone’s life story and you get competent, fluent, generic prose. It will fill gaps you did not tell it about. It will smooth a specific memory into a familiar shape. It will produce something that reads well and is not true, and the person whose life it is will not be able to say precisely what is wrong — only that it is not theirs.

Both halves of that failure have the same root: the model is being asked to write before it has anything to write from. So the workflow makes that structurally impossible.

Interview before drafting, always

Drafting cannot begin until an interview has happened. The interview turns a one-sentence memory into a scene, and it asks for things a model cannot invent:

• Sensory anchors — colour, smell, sound, temperature, what was eaten, what it cost, what brand it was
• Actual words — who said what, in which dialect, phrased how
• Gesture — what were their hands doing; a characteristic tic or catchphrase
• Place and time — season, weather, room layout, the route taken
• The gap — what you believed then against what you understood later
• Where it ended up — what became of that object, that person. The afterwards is very often the ending.

Two rules govern the transcript. Spoken answers are recorded close to verbatim — the roughness of speech is the raw material, and “tidying it up” at the recording stage destroys the thing you came for. And “I don’t remember” is recorded as “I don’t remember”, tagged as such. A memoir is allowed to say the author has forgotten. It is not allowed to make something up.

The style profile, and why it improves

Voice is defined by a separate specification derived from the author’s own writing across 2014–2026 — not a description of good prose, but a description of this prose: its habits, its tells, its failure modes.

It carries explicit prohibitions, and they are the useful part because each names a specific way this author’s writing goes wrong under pressure:

• End on the thing that happened, not on what it means
• No essay register — none of the throat-clearing that precedes an argument
• Empty adjectives get replaced by a noun or an action
• At most one abstract noun per chapter
• Loanwords capped, counted

The mechanism that makes the system improve is the feedback loop. After every editing pass, what the author changed and why is distilled into one to three new rules and appended to the style profile. Not a note about the chapter — a rule for the next one. This step is marked as non-skippable, because it is the only reason the output converges on the author’s voice instead of hovering near a general-purpose one forever.

That is the transferable idea: a style guide that is written once decays, because the interesting corrections are the ones you did not anticipate. A style guide that absorbs each round of edits becomes a genuine model of a specific writer, and it does it using the cheapest possible signal — the diff between what was drafted and what was kept.

Provenance as a hard constraint

Every detail in a finished chapter must trace to either an interview transcript or a document in the source corpus. When a detail is missing, the instruction is not to write around it — it is to go back and ask.

This is slower. It is also the only version of this tool worth having. A memoir is a claim about what happened, made to the people in it, and a system that produces plausible filler produces something the author cannot honestly sign.

Why there is nothing to show

The output is a private autobiography and will not be published — not in excerpt, not as a sample. What is portable is the shape: interview → draft under an explicit voice spec → edit → distil the edit back into the spec.

I have since reused that pattern for work with nothing to do with memoir. Any task where quality means matching a particular person’s judgement rather than meeting a general standard has the same structure, and the same answer: stop describing the target, and start harvesting the corrections.