conorbronsdon/avoid-ai-writing

Surfaced by CodeRabbit round 2 on wshobson/agents#645: edit mode takes --file PATH with no statement about what kinds of files are in scope. Pointed at source code, configuration, or generated data, a prose rewrite corrupts the file. CodeRabbit proposed a Markdown-only gate; the port declined that shape because the skill explicitly covers plain-text prose registers (the curly-quotes rule names code comments, commit messages, and plaintext drafts), and went with prose-vs-structured instead: "The target is a prose file: refuse source code, configuration, and generated data, and say why." (wshobson/agents#645 commit ed69ec8) SKILL.md §edit mode here has the same gap. Same fix shape should work, worded to fit the fuller edit-mode paragraph.

Surfaced by CodeRabbit on the wshobson/agents port (wshobson/agents#645): three voice-profile targets mandate content the source may not contain, which collides with the "Never inject these" guardrails when the source lacks the target. casual: "At least one first-person or concrete-anecdote touch" vs "Fake first person: ... If the source has no I, the rewrite has no I." professional: "One concrete claim per paragraph (a number, a name, a date)" and "Make the ask explicit" vs "Invented specifics: ... If the concrete detail is missing, flag the gap and leave it. Never fill it." warm: "acknowledge them at least once" can require inventing an acknowledgment. The Never-inject section wins by its own terms ("None of the following may be added"), but the profiles never say so, and a model following the targets checklist-style will fabricate the missing touch. The port resolved it by bounding each target on source evidence, e.g. casual became "Keep first-person and concrete touches where the

Research for #88 turned up a gap worth taking. Vale is a prose linter with 5,712 stars, a package registry, a GitHub Action, editor plugins for VS Code/Atom/Sublime, and a language server. Its registry ships 16 packages: AsciiDoc Elastic Google Hugo Joblint MDX Microsoft NoAnimalViolence OpenShiftAsciiDoc Readability RedHat Salesforce alex neighbor proselint write-good None of them targets AI-writing tells. proselint, write-good, alex and Readability cover prose quality, insensitivity, and grade level. Nothing covers what this repo catalogues. Why this is the right shape for distribution The audience already exists, with CI and editor integration built. No new schema for anyone to adopt. Zero new guide content to maintain. The rules are already ours. The detector is regex-based, which is most of the work of a Vale style. It would be the only package in that registry with a measured false-positive rate rather than an asserted one. Every other one is asserted. That is the

Split out of conorbronsdon/personal-context#119 (Tier 3), which is now closed so this has one home. What exists and what doesn't detector/ is a working Node CLI with tests, and .github/workflows/detector-test.yml already runs npm run self-scan:check against this repo's own docs. So the detector is proven as a CI gate. It just isn't packaged for anyone else to use as one. There is no action.yml and no .pre-commit-hooks.yaml. A user who wants the detector on their own docs has to wire the CLI themselves. The two artifacts GitHub Action. A composite action taking a glob, a threshold, and a context profile, failing the check when a file exceeds threshold. The self-scan workflow is the working reference for what the gate looks like; the work is generalising it beyond this repo's paths. Prior art: Aboudjem/humanizer-skill. Pre-commit hook. A .pre-commit-hooks.yaml entry so pre-commit users get the detector on staged files. Prior art: patina's .pre-commit-hooks.yaml. Why this is the safest Ti

Split out of conorbronsdon/personal-context#119, which is now closed. Documented in corpus/README.md since the corpus landed, but never tracked as work. The gap corpus/manifest.json covers, on the human leg: Register Entries blog 26 essay-literary 4 academic 3 mixed 2 technical-blog 1 Zero entries for docs, social, email, and conversational — four of the registers the tolerance matrix defines and the tool is pointed at every day. LinkedIn posts and release notes are the actual use case and neither is represented. Why this blocks more than it looks like corpus/README.md states the gate and then fails it: no number from this corpus goes into the README, a release note, or a social post until each claim cell has n >= 100, covers more than one register that people actually write in today, and carries a confidence interval. The current run satisfies the interval and the n, and fails the register test outright. So the measured 0.0% FP rate cannot be published. Patina's