Field manual for AI-assisted work
Cited rules for the defaults that make unrelated AI work look the same.
Or install it as a skill: npx skills add jtc268/not-like-us. Sources and full guides are on GitHub.
Use the Not Like Us rules for every writing and interface decision.
Rules: https://raw.githubusercontent.com/jtc268/not-like-us/main/AGENTS.md
Start from the audience, task, real content, and existing brand. Local design systems and voice guides win.
Reject visual choices the tool supplied without a reason: purple gradients, aurora blobs, glass panels,
rounded cards around everything, Inter or Geist by reflex, centered hero formulas, three equal columns,
decorative pills and icon tiles, fake charts and testimonials, filler imagery, and motion that explains nothing.
Reject empty writing patterns: setup before the point, "not X but Y" slogans, vague consensus, hype adjectives,
rule-of-three filler, synonym cycling, stacked fragments, recap endings, and em dashes.
Use real data and real states. Name a source for every factual claim. Preserve the author's voice.
Never claim that one tell proves AI authorship.
If the work could belong to any product after swapping the logo, make it specific.Filter by medium or generator. Each row says what to reject and what to try next.
Separate writing and design checks for each generator.
Multiple design directions, with a separate standard dashboard flow.
Documented cream, serif, italic-accent, and terracotta house style.
Explicit frontend guidance rejects common decorative and card-heavy defaults.
Broad generation needs a narrow voice and visual brief.
shadcn/ui, Next.js, and Tailwind are documented defaults.
Fast full-stack generation can spread early defaults through the build.
Existing repository constraints are the best defense against visual drift.
Design Mode starts front-end-only, then converts into an application.
Themes and global AI design changes need careful scoping.
Host-tool defaults and model choices need separate inspection.
Auto-layout cards can shape the argument before the evidence does.
This page is a snapshot. The Stream keeps your agents on the current rules without a reinstall, in Claude Code, Codex, Cursor, OpenClaw, Hermes, Gemini CLI, and Copilot.
Checking the stream version.
A script fetches every source and records what changed. People approve the guidance.