# Agent-first Writing Editor pitch

> A repository-first path from deterministic prose checks to evidence-backed editorial assistance

- HTML version: https://robbiepalmer.me/projects/agent-first-writing/deck
- Project: Agent-first Writing Editor (https://robbiepalmer.me/projects/agent-first-writing.md)
- Slides: 8

## Slide 1

# Good research still deserves a readable ending

Agent-first Writing Editor tests how to turn dense agent output into prose people can absorb without erasing the author's voice.

## Slide 2

## Correct sentences can still exhaust the reader

Agent drafts often repeat conclusions, flatten the author's voice, and explain simple points long after the point is clear.

Grammar is part of the problem. Editorial judgment is the harder part.

## Slide 3

## Vale is the working first layer

### Shipped

Repository rules flag known prose faults at exact source locations before review.

### Bounded

A finding identifies a problem. It does not invent a replacement or claim that an edit was accepted.

Use a fast rule when the rule can prove the fault.

## Slide 4

## Detection and action need different records

### Finding

A verified source span needs attention, but no safe replacement exists.

### Suggestion

One actionable replacement for a verified source span.

### Proposal

One or more dependent suggestions shown as a single change.

### Decision

The author accepted, rejected, or changed the complete proposal.

## Slide 5

## Grammar correction did not finish the edit

### The test

Run GECToR-2024 against existing architecture decisions as a lightweight correction baseline.

### The result

Some corrections helped. The prose remained repetitive, over-explained, and harder to absorb than the published version.

The project rejected GECToR as its model layer and kept the result as a lower bound.

## Slide 6

## Intervention should grow with editorial risk

1. **Current.** Exact Vale findings flag known faults.
2. **Foundation built.** Shared records can hold local suggestions and dependent proposals.
3. **Planned next.** Models target only failure classes that the evaluation set proves rules miss.
4. **Planned later.** Passage rewrites and narrow automatic fixes follow reliable local suggestions.
5. **Human authority.** A person reviews work published under somebody's name.

## Slide 7

## The repository is the proving ground

```mermaid
flowchart LR
A[Pinned draft and published revisions] --> B[Frozen evaluation cohort]
B --> C[Vale and baseline producers]
C --> D[Verified findings and proposals]
D --> E[Observed edits]
E --> F[Explicit editorial decisions]
F --> G[Scorecard and release decision]
```

Git proves that text changed. Only an explicit decision proves what the author accepted.

## Slide 8

## Real use must prove the product boundary

Before extraction, the repository release must show that edits preserve facts, intent, terminology, and voice while reducing review time and manual rewriting.

It also needs recorded accept, reject, and change decisions, plus a reproducible baseline for every candidate producer.

Success means readers prefer the edited work and authors do less rewriting.

---

Markdown index of this site: https://robbiepalmer.me/llms.txt