https://github.com/0plus1/non-fiction-workbench

https://github.com/0plus1/non-fiction-workbench

Last synced: about 1 month ago
JSON representation

• Host: GitHub
• URL: https://github.com/0plus1/non-fiction-workbench
• Owner: 0plus1
• License: mit
• Created: 2026-04-18T04:46:26.000Z (2 months ago)
• Default Branch: main
• Last Pushed: 2026-04-18T05:25:11.000Z (2 months ago)
• Last Synced: 2026-04-18T07:25:45.726Z (2 months ago)
• Size: 13.7 KB
• Stars: 0
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# Non-Fiction Workbench

LLM agent skills for long-form non-fiction writing and revision.

- `/non-fiction-workbench` for drafting, rewriting, and critiquing non-fiction against a brief, source set, and house style.

- `/anti-slop` for detecting and removing common LLM prose tells without changing factual meaning.

## Why This Exists

AI-assisted non-fiction usually fails in predictable ways:

- claims get stronger than the evidence

- structure drifts away from the brief

- paragraphs bloat with transitions, summary, and throat-clearing

- rewrites flatten voice and introduce generic "smart" phrasing

This repo gives agents a tighter workflow for essays, articles, reported features, newsletters, and book chapters. It does not replace authorship or reporting. It gives the model a disciplined way to read the draft, inspect the brief and source material, and revise without hallucinating or padding.

The anti-slop skill is informed by common LLM prose-tell catalogs, including the public taxonomy at [tropes.fyi](https://tropes.fyi/tropes-md), then distilled into rewrite-oriented rules for actual editorial use. It also includes a stricter parity checklist for detector-style passes, based on the rule inventory used in [Slop Cop](https://github.com/awnist/slop-cop).

## What's Included

```text

non-fiction-workbench/

├── README.md

├── LICENSE

└── skills/

    ├── non-fiction-workbench/

    │   ├── SKILL.md

    │   ├── agents/

    │   │   └── openai.yaml

    │   └── references/

    │       └── project-shape.md

    └── anti-slop/

        ├── SKILL.md

        ├── agents/

        │   └── openai.yaml

        └── references/

            └── slop-patterns.md

```

## Install

Install the core writing skill:

```bash

npx skills add https://github.com/0plus1/non-fiction-workbench --skill non-fiction-workbench

```

Install the anti-slop skill:

```bash

npx skills add https://github.com/0plus1/non-fiction-workbench --skill anti-slop

```

## Usage

Invoke the core skill manually:

```text

/non-fiction-workbench write drafts/article.md

/non-fiction-workbench rewrite drafts/article.md

/non-fiction-workbench critique drafts/article.md

/non-fiction-workbench global-critique drafts/

/non-fiction-workbench global-rewrite drafts/

```

Invoke the anti-slop skill manually:

```text

/anti-slop scan drafts/article.md

/anti-slop revise drafts/article.md

/anti-slop global-scan drafts/

/anti-slop global-revise drafts/

```

## Modes

`non-fiction-workbench`

- `write`: draft or expand a non-fiction piece from the brief, outline, and source material

- `rewrite`: improve an existing draft while preserving thesis, facts, and structure

- `critique`: return editorial feedback on argument, evidence, structure, and prose

- `global-critique`: critique a whole draft directory as one manuscript

- `global-rewrite`: rewrite a whole draft directory while preserving argument and evidence

`anti-slop`

- `scan`: flag clear AI-style rhetoric and document-level repetition

- `revise`: rewrite to remove AI tells while preserving factual content and argument

- `global-scan`: scan a full draft directory for repeated AI-style patterns across the manuscript

- `global-revise`: remove slop across a full draft directory while preserving meaning

## Expected Project Shape

The skills are opinionated about project structure. They work best when your project looks roughly like this:

```text

your-book-or-essay/

├── brief/

│   ├── thesis.md

│   ├── audience.md

│   ├── constraints.md

│   └── outline.md

├── research/

│   ├── source-index.md

│   ├── notes/

│   └── excerpts/

├── style/

│   ├── voice.md

│   └── house-style.md

└── drafts/

    ├── chapter-1.md

    ├── essay-1.md

    └── article-x.md

```

The core skill reads the target draft first, then consults the relevant brief, style, and source files. For manuscript-level work, it can read an entire `drafts/` directory in order. The anti-slop skill reads the target draft and applies a conservative rule set for common LLM prose patterns, including repeated rhetorical scaffolds that become obvious only at manuscript scale.

## How They Behave

Both skills are deliberately manual. They use `disable-model-invocation: true`.

They tell agents to:

- read the target draft first

- inspect the relevant brief, style, and source files before changing claims or structure

- preserve the author's thesis, evidence, and level of certainty

- avoid inventing facts, sources, examples, chronology, or causal claims

- cut generic LLM phrasing, padding, and rhetorical scaffolding

- keep the human in charge of judgment, reporting, and final wording

For `write` and `rewrite`, the core skill is intended to revise the target draft directly when appropriate. For `critique` and `scan`, it should usually return notes unless you explicitly ask it to write those notes into project files.

## Philosophy

This repo is not an autopilot content generator.

The author still owns:

- thesis

- reporting

- structure

- taste

- final language

The agents' role here is narrower:

- reduce factual drift

- keep the brief and source material in view

- make revision more disciplined

- remove generic AI prose habits before publication