# Substack Writing Harness

> Substack Writing Harness is an AI agent skill for Claude Code, Codex CLI, Gemini CLI, and other agents that read SKILL.md. Triage the idea, draft the essay, and push the score to zero before you publish. Best for: AI Substack writing.

- URL: https://atharvashah.com/ai-skills/substack-writing-harness
- Price: $49, one payment
- Works with: Claude Code, Codex CLI, Gemini CLI, Cursor, GitHub Copilot
- Category: Writing and voice
- Built by: Atharva Shah (https://atharvashah.com)
- Requirements: Needs a paid Claude plan with Claude Code and Python 3. Triage needs your published posts as local text or Markdown files. Publishing stays with you, since the harness stops at a publish-ready draft.
- Buy: https://buy.polar.sh/polar_cl_R6zXD453Fqls2A2MvMzI0peVQkdv2aw1TZP2c1jBqZo?custom_field_data.skill-premium=substack-writing-harness (the skill folder and setup guide arrive by email)

## Write the essay only you could write, then prove it.

It checks your idea against everything you already published, pulls your real opinion out before a word is drafted, and runs the draft through scorers until every blocking finding is gone.

## What it produces

- **The published essay.** You publish an essay that opens on its claim and on work only you did. ![A real published Substack article with its cover and an opening paragraph that states the claim.](https://atharvashah.com/ai-skills/substack-writing-harness/scene-1.webp)
- **The blocking checks.** Every scorer reaches zero before a draft is called ready. ![A real polish report: 131 drafts at zero blocking findings across three scorers.](https://atharvashah.com/ai-skills/substack-writing-harness/scene-2.webp)
- **Questions, not guesses.** Facts it cannot verify come back to you as questions. ![A real list of open questions the harness raised about drafts instead of guessing.](https://atharvashah.com/ai-skills/substack-writing-harness/scene-3.webp)

## A fixed loop, so no step gets skipped.

"Write a post about X" and "polish this" both mean the whole loop.

1. **Triage the idea.** It measures how much of the idea already lives in your published posts. A covered idea becomes a section of the old post, which usually makes both stronger.
2. **Intake, then draft.** It asks for your hot take, the proof only you have, and the reader. Then it drafts on a four-beat arc and cuts ten percent before any scorer runs.
3. **Score, polish, ship.** Scorers run until blocking findings reach zero. Then it checks every image, embed, and outside fact, and a second reader looks for contradictions.

## AI writing tools produce the average post in your niche.

Your readers can find the average post anywhere. They subscribe for your opinion.

### Substack Writing Harness

- **Archive-aware triage.** It checks the idea against every post you published.
- **Opinion first.** Intake pulls your take out before a word is drafted.
- **Scored to zero.** Slop, clarity, and craft checks all reach zero.
- **A real polish pass.** Images, embeds, diagrams, and outside facts get checked.

### AI writing apps

- **Opinion-free drafts.** They write around a topic and commit to nothing.
- **Blind to your archive.** They happily write your fourth post on the same idea.
- **The house voice.** Every draft carries the same tells readers now skip.
- **Unchecked facts.** Claims about other people ship without a source.

### A human editor

- **Days per round.** Your timely post goes out a week late.
- **Your archive is a blur.** They cannot recall what you wrote two years ago.
- **Fact checks cost extra.** Most line edits skip them.
- **A bill per issue.** Weekly publishing makes it a second rent payment.

## Frequently asked questions

### Does Substack allow AI written content?

Substack allows it, and readers are quick to spot it. This skill keeps your argument and your facts in front and removes the patterns readers flag as AI before the post goes out.

### How do I spot AI writing on Substack?

Common tells are restated headings, three-part lists, and claims with no source. The skill checks for those patterns by name and fixes each one in your draft.

### Does it only work with Substack?

It was built for Substack, and it works for any long-form newsletter or blog. The drafts are Markdown files you paste or upload.

### How does triage know my old posts?

You export or save your published posts as local files once. The overlap check reads that folder each time.

### Will it write my opinion for me?

No. Intake asks for your take and your proof. If the answers are soft, it pushes once and then waits.

### What if I want to write the idea anyway?

It pushes back once, then writes it. It records your override so a later check does not argue again.

### How long are the essays?

800 to 2,000 words, with 1,800 as the target. It tells you when an idea is honestly shorter.

### Which Claude plan do I need?

A paid plan with Claude Code. Pro works for one essay per session, and a full polish pass on a long piece can hit limits.

### What else do I pay for?

Nothing. The scorers are local Python scripts. Diagram and image checks use free tools.

### Does it publish for me?

No. It stops at a publish-ready file and a checklist. You press publish.

### What if the first draft misses my voice?

Paste three past essays you like into the setup. The drafts match that rhythm, and every essay you finish adds to the archive triage reads.

### Does this skill work with Gemini CLI, Codex, Cursor, or GitHub Copilot?

Yes, with one caveat. The skill is a SKILL.md folder in the Agent Skills open standard, which Claude Code, Codex CLI, Gemini CLI, Cursor, and GitHub Copilot all load. It is built and tested on Claude Code. Steps that lean on Claude-only features, such as parallel subagents, run one after another or need a manual step in other agents.

### How do I install an agent skill?

Copy the skill folder into your agent's skills directory: ~/.claude/skills for Claude Code, or the skills folder your agent documents for Codex CLI, Gemini CLI, Cursor, or Copilot. The agent reads the description and loads the skill when your request matches it. Each skill ships with a setup guide for any tool it needs.

[All AI skills](https://atharvashah.com/skill-pack) · [Library index](https://atharvashah.com/ai-skills.md)
