Verbatim

Verbatim

An engine and a profile. The engine ships in this repository. The profile is one person's material and never leaves their machine.

The promise, and the constraint that produces it: it cannot write anything you did not say. Every fact in a generated post is traceable to the profile, to the corpus, or to a sentence spoken in the interview that produced it. When nothing traces, nothing gets written.

Route

Read references/ on demand, not up front. Each skill names the files it needs.

This router names almost no paths, deliberately. It travels inside every generated engine, and an engine carries every path it names, so a see-also here costs its file six times over. Where a sentence can carry the fact, it does.

First thing, every time: the status flag

Read the ## Status block at the top of the material the person attached:

## Status
- filled: no
- source: template
- updated: --
- interview_language: --
- output_language_default: --

While filled: no, no skill pretends to know the person. Fall back to generic rules, say so in one line, and offer linkedin-setup. This three line block is the whole seam between the engine and a profile.

The profile directory is whatever the person points at. linkedin-setup creates it and writes its path into the conversation. There is no default path in this repository, on purpose: a hard coded path is how an engine stops being portable.

The three language axes

Independent, and this is what multilingual projects usually get wrong.

The last two really are independent. Plenty of people want to be interviewed in their own language and publish in English.

Language leak is the failure mode to watch. If a template in this engine carries a label like FIRST LINE, a model will echo it in English inside a French post. Two guards, both mandatory:

1. No user-facing string lives in the engine. Wording lives in locales/<lang>/interview.md. The engine holds intents, never sentences to say.
2. Every generation opens with an explicit output-language directive. State it, in one line, before writing anything.

If the pack has no wording for an intent, generate one and announce it: degradation is visible, never silent.

Language packs

locales/en and locales/fr ship. locales/_template is the contract, and its README carries the acceptance criteria for a new one.

A pack is never a translation of another pack. The ten style categories are shared across every pack; the word lists that fill them are not. The same word is a marketing tell in one language and ordinary in another.

Layout

SKILL.md                    this router
skills/linkedin-post/       interview, sheet, draft, check, publish, measure
skills/linkedin-setup/      onboarding, pillars, idea bank
skills/linkedin-profile/    the public page, nine sections, audit then rewrite
references/                 mechanism, shared across languages
locales/<lang>/             wording, style lists, market rules
lib/lint.py                 deterministic style pass, no model
lib/publish.py              the three publishing tiers
examples/                   a fictional persona, to read before running this

What this bundle will not do

• No hook formulas calibrated on a viral corpus. They invert the mechanism: the angle would descend from a template instead of from something the person said.
• No writing against an AI detector. Optimising for a classifier is writing for the classifier. The deterministic pass in lib/lint.py is the whole of it.
• No engagement pods, and no comment gate by default. A comment gate is a deliberate choice made once there are enough published posts to know whether the audience plays that game, never a default.
• No invented facts, ever, including in a revision. That is the promise.