"My Editor Caught Me Sounding Like AI. Now AI Catches Me First." — Katie Parrott

Why this is in the vault

A working writer codifies her editor's catch-list into a named review skill (/guardrails) plus a roster of persona reviewers, which is a near-exact mirror of RDCO's skills-as-craft and voice-match draft pipeline.

The core argument

Parrott (Every's "Working Overtime" columnist, the staff writer who writes about AI-assisted writing using AI) got flagged by editor-in-chief Kate Lee for drafts that had absorbed the classic AI tells: symmetrical sentence structures, rhetorical throat-clearing, hollow profundity, and rule-of-three padding. Her response was not to write more carefully by willpower but to bake the catch-list into documentation for her agents: a Markdown skill she calls /guardrails, which turns any agent into an exacting editorial reviewer.

Her central reframe leans on Dan Shipper's "After Automation" thesis: AI does not reduce the writer's workload, it changes what the work is. To make a model enforce her standards she has to define those standards explicitly, which is more work but forces her to understand her own preferences better. Articulating the rules "wears new grooves" into her brain.

She runs a staged gauntlet, not a single pass: outline stage gets adversarial structural reviewers (Hitchcock for tension, Sedaris for humor, /asshole for the least-charitable read of the argument); section drafts get /ai-check and /guardrails to catch local smoothing; the full draft gets a developmental review plus a column-specific skill (/working-overtime) that checks her signature structural moves; a final pass reruns /ai-check, /guardrails, /tighten-draft, and /kate-top-edit (built around her editor's specific tics: vague "this" openers, unsourced data, floating quotes, hedges, missing internal links).

Two craft points worth keeping. First, naming matters: she gives editorial skills personas (Sorkin keeps the piece "walking and talking," Mom flags where a non-AI-native reader gets lost) because clinical names like "assess narrative momentum" don't stick, though she notes this matters less as models get better at tool selection. Second, orchestration: a /panel command convenes multiple reviewers in parallel, then a synthesizer reads their feedback together, preserving productive disagreement (one reviewer wants a section cut, another calls it the most alive thing in the draft) because the tension surfaces the actual decision the writer has to make. The reviewers are explicitly a pre-filter so the human editor's attention is reserved for what only a human can do: angle, claim, storytelling, audience fit. She closes by pairing the piece with an Every GitHub repo of the reviewer skills, framed as an example to adapt, not a blueprint to copy, since the value is in setting and enforcing your own standards.

Mapping against Ray Data Co

Strong. This is the highest-fidelity external validation yet of the exact pattern RDCO already runs.

• Review-skill-as-editor maps to RDCO's existing skills. /draft-review and /voice-match are functionally Parrott's /guardrails + /kate-top-edit: a codified catch-list that pre-filters a Sanity Check draft before founder/human attention. The "fresh-eyes subagent for own-artifact review" memory (/design-critic, /video-critic, /verify-*) is the same insight applied to non-prose surfaces, and /panel-with-synthesizer is the multi-axis fan-out RDCO already uses in pipeline-critic and the verification-as-independent-worker SOP.
• "Articulating standards for the model wears grooves in your brain" is the founder's own thesis, externally confirmed. Building the Copilot Studio agent-builder plugin and the skills library is itself the act of making tacit judgment explicit. Worth a Sanity Check angle, but per the no-derivative rule it needs an original re-frame, not a restatement; the non-obvious move here is that the review skill is an editor you author, so the bottleneck shifts from "write well" to "specify taste precisely."
• Persona-naming for stickiness validates RDCO's named-critic convention (/design-critic, /video-critic) over clinical task names, with the honest caveat that the benefit decays as tool-selection improves in newer models.
• Same author, deepening thread. Parrott also wrote the Codex-for-knowledge-work and editing-AI-writing pieces already in the vault; this is a tracked author building a consistent "editing/review is the new writing skill" position RDCO's content pipeline is built around.

⚠️ Bias / self-promotion (no third-party sponsor)

No paid issue sponsor. The email carries a generic "for sponsorship opportunities, reach out to sponsorships@every.to" footer (solicitation, not a placed ad) and an Every product-suite house promo (Sparkle, Cora, Spiral, Monologue) plus a subscription upsell. The piece also self-distributes Every's own GitHub reviewer-skills repo (no affiliate/tracked monetization). Tooling mentioned (Codex, Claude Code) is illustrative, not sponsored.

Related

• [[2026-05-13-writewithai-cowork-skill-for-fiction-voice]]
• [[2026-03-18-every-editing-ai-writing]]
• [[2026-04-26-write-with-ai-build-claude-skill-from-scratch]]
• [[2026-05-26-every-codex-for-knowledge-work-power-user-guide]]