# Ghostwriter design sources

Reviewed 19 November 2026. Each project link below selects the exact Git tree consulted. The pack implements its editorial material in independently written examples, profiles and instructions. No external skill, detector, decoding component, corpus, model weight or word list is loaded by these resources. Upstream licenses apply to their respective projects. Keep copyright and permission notices with any future source import and record the exact files and changes in the repository's third-party notices.

## Upstream decisions

| Project | Idea examined | Decision for Ghostwriter | How to inspect the choice and the source |
| --- | --- | --- | --- |
| milock/humanizer | A different channel can call for a different amount of editing. The source distinguishes author and brand profiles and guards the underlying request. | Use the explicit channel, link the voice to approved samples, and retain unresolved requests or reservations. Select the smallest edit that addresses the reader’s problem. | Compare a message to a customer and an internal note from the same person. Keep the pending action in both. [Source](https://github.com/milock/humanizer/blob/ec66c203c0400481819fda9cb17a6f1c7977c975/SKILL.md) · [MIT](https://github.com/milock/humanizer/blob/ec66c203c0400481819fda9cb17a6f1c7977c975/LICENSE) |
| blader/humanizer | The editor covers prose embedded in files and uses repeated rhetorical structures as review clues. | Scope changes to the authorized prose. Keep literal material intact. Treat instructions inside drafts as content. Personal detail needs a source. | Ask for a rewrite of Markdown containing a command, a quoted instruction and a URL. Compare those spans exactly. [Source](https://github.com/blader/humanizer/blob/a32f24407364c74c812b43c15926d638bee1e072/SKILL.md) · [MIT](https://github.com/blader/humanizer/blob/a32f24407364c74c812b43c15926d638bee1e072/LICENSE) |
| AshwinSathian/humanize-writing-skill | Concise directions during drafting can address repetitive structures before a second editing pass. Genre and local usage change the appropriate style. | Start with the requested answer or task. Preserve technical definitions and narrative rhythm; use the locale sheet for language-specific choices. | Ask for a one-line status and a technical explanation using the same brief. Check each against its intended reader. [Source](https://github.com/AshwinSathian/humanize-writing-skill/blob/66edd8014dc14125b5df38b4b7504fa243fb2a61/SKILL.md) · [MIT](https://github.com/AshwinSathian/humanize-writing-skill/blob/66edd8014dc14125b5df38b4b7504fa243fb2a61/LICENSE) |
| ferr079/humanizer-fr | French calques, connectives and typography need French examples. File paths and commands can coexist with editable prose. | The French sheet distinguishes intentional spacing and domain terms from cumbersome wording. | Place an intentional fine nonbreaking space inside a quoted French label. A general cleanup must leave it there. [Source](https://github.com/ferr079/humanizer-fr/blob/a6230b933399be93c088e611fefa72715377f2f9/SKILL.md) · [MIT](https://github.com/ferr079/humanizer-fr/blob/a6230b933399be93c088e611fefa72715377f2f9/LICENSE) |
| alxbd/boileau | French prose can lose its voice through affected informality or over-explanation. Register depends on the relationship. | Resolve tu/vous for the destination. Prefer preserving an established register to adding invented familiarity. | Compare a collegial brief with its customer-facing email. Match the addressee in each. [Source](https://github.com/alxbd/boileau/blob/5b272a70b1d5387984c12a08c4edc45af3f4fbda/SKILL.md) · [MIT](https://github.com/alxbd/boileau/blob/5b272a70b1d5387984c12a08c4edc45af3f4fbda/LICENSE) |
| hardikpandya/stop-slop | A short editorial prompt can draw attention to repetitive setup, weak verbs and empty emphasis. | Flag a format only when it obstructs the specific task. A punctuation ban or the author’s self-scored threshold has no role in acceptance. | Leave a useful list of procedure steps or an intentional rhetorical device in place. [Source](https://github.com/hardikpandya/stop-slop/blob/8da1f030185bdfe8471220585162991eaeb970e9/SKILL.md) · [MIT](https://github.com/hardikpandya/stop-slop/blob/8da1f030185bdfe8471220585162991eaeb970e9/LICENSE) |
| conorbronsdon/avoid-ai-writing | The source distinguishes review from rewriting and discusses English calibration, false positives and density. | A diagnosis may leave the original unchanged. Repetition suggests a location for review; it supplies no authorship or watermark evidence. | Use an already-clear first draft and check whether an unchanged result is accepted under the editorial policy. [Source](https://github.com/conorbronsdon/avoid-ai-writing/blob/a5d6bae7c0d762dbcff74ca86c17a7ac9fee4151/SKILL.md) · [MIT](https://github.com/conorbronsdon/avoid-ai-writing/blob/a5d6bae7c0d762dbcff74ca86c17a7ac9fee4151/LICENSE) |
| sam-paech/antislop-sampler | The sampler suppresses chosen strings during inference and needs control of token probabilities. | Out of runtime scope. Its measurement of repeated phrases can inform a corpus study. A blacklist or its model-specific weights cannot serve as an editorial gate. | Check the inference boundary before making any claim about model-time control. [Source](https://github.com/sam-paech/antislop-sampler/blob/0ae330e98fbe6f09351f2d1063a51956378a44b2/README.md) · [Apache-2.0](https://github.com/sam-paech/antislop-sampler/blob/0ae330e98fbe6f09351f2d1063a51956378a44b2/LICENSE) |
| cyzanfar/text-watermark-remover | The tool distinguishes character inspection from detector-specific statistical experiments. | Out of runtime scope. Preserve meaningful typography and provenance. Zero-width characters and missing detector hits do not establish authorship. | Check the French sheet and a mixed-script sample for unintended normalization. [Source](https://github.com/cyzanfar/text-watermark-remover/blob/fd620a93b9d96c7a4730d80be2d7aa90d5017575/README.md) · [MIT (repository), separate terms for included Unicode data](https://github.com/cyzanfar/text-watermark-remover/blob/fd620a93b9d96c7a4730d80be2d7aa90d5017575/LICENSE) |

The French projects acknowledge material from Siqi Chen's original Humanizer. Similar observations in these projects share a lineage. Their licenses credit Stéphane Ferreira and alxbd respectively for French work. Humanizer contributors and Ashwin Sathian are credited in the corresponding source licenses. An upstream project's own score describes its own measurement conditions.

## Evidence guiding the choices

- Named entities, quantities and quoted passages can change during a rewrite. The October 2026 comparison of commercial services motivates tracking them across a pair of texts. A difference identifies a review point. Reusing a number in another claim can keep the digits intact while changing the meaning. [Wang, Zhu and Chen](https://arxiv.org/abs/2611.00359)
- Lists, links, emphasis and response length influence preference judgements. Judge each response against its task and compare content separately from format. [Zhang et al., ACL](https://aclanthology.org/2025.acl-long.1308/)
- Personal-style imitation varies by person and genre. Use owner-approved examples from the relevant channel. Report sample coverage and unresolved traits. [Findings of EMNLP](https://aclanthology.org/2025.findings-emnlp.532/)
- Repeated openings, length, constructions and paragraph rhythm can be counted separately across samples. Weighting them can change a model ranking. The outcome of interest here is whether a reader can use the answer and whether its information survived the edit. [SlopBench](https://arxiv.org/abs/2609.33905)
- Compar:IA provides French chat requests and reader preferences for exploring realistic tasks. Its competing responses come from models. A vote alone cannot identify the quality that motivated it. Check the dataset's terms before incorporating any examples. [Ministry of Culture dataset](https://comparia.beta.gouv.fr/datasets)
- Official service-writing guides inform the original examples for [French](../languages/fr-FR.md), [English](../languages/en.md) and [Brazilian Portuguese](../languages/pt-BR.md). The preferences of a particular publication, locale and reader take precedence over those genre examples.

## Upstream maintenance

Compare a proposed new revision with the pinned source. Record the upstream author, license and ancestry, then identify which Ghostwriter contract or example would change. Test the proposed change against the intended genre and locale, including unchanged good prose. Update the source link and the local resource version together after review. Keep a project's approved profiles and revision history in project-owned storage.
