How to Make AI Writing Sound Human with Agent Skills

Do not give an AI one vague instruction to “sound human.” Give it a small, testable writing system. Separate source knowledge, response structure and editorial quality. Then measure whether each layer improves the work. That is the practical lesson behind three fast-growing open-source projects: No AI Slop, i-have-adhd and book-to-skill.

They are often presented as three ways to fix the same model. They are not. No AI Slop edits recurring prose patterns. i-have-adhd changes how an assistant structures an answer. book-to-skill turns a large reference set into knowledge the agent loads when relevant. Used together, they can reduce repetitive cleanup. Used without source material, evals or human judgment, they can produce tidy writing that is still empty or wrong.

This article owns the writing-workflow question. If you first need to understand the technology category, our MCP vs RAG vs Agent Skills vs Custom GPTs decision tree explains where skills sit in an AI system.

The short verdict: which AI writing skill solves what?

For one person, install the narrowest skill and keep editing. For a company, treat the skill as a versioned product component. Define what it may change, how success is scored, who approves claims and how a bad update is rolled back.

Why does AI writing sound generic?

Most weak output starts before the sentence is written. The model receives an underspecified task, little first-hand material and no acceptance test. It then optimizes for a plausible average answer. The result is smooth structure, broad claims, repeated rhetorical patterns and conclusions that sound more important than the evidence supports.

A longer prompt is not automatically better. Mixing facts, tone, process, forbidden phrases, audience, formatting and examples into one permanent block makes conflicts harder to see. It also turns every change into prompt surgery. A reusable skill gives those instructions an owner, a version and a place for examples or deterministic checks. The related concept is prompt engineering, but the operational improvement comes from packaging and testing the procedure, not finding one magical sentence.

Three open-source skills, three different jobs

1. No AI Slop: a focused post-editor

No AI Slop documents more than 20 recurring patterns, including throat-clearing openings, fake-insight setups, dramatic fragments, vague attribution and inflated conclusions. It can edit a draft or identify matching lines without claiming to detect who or what wrote the text. The repository includes the skill instructions and an evaluation file, which makes its editorial rules inspectable.

That narrow scope is its strength. Run it after the factual draft exists. Do not let a style pass invent proof, strip necessary qualifications or force every author into one cadence. “Human” is not a style preset. The useful goal is specific writing whose source, argument and voice remain intact.

2. i-have-adhd: action-first response structure

i-have-adhd tells a coding assistant to lead with the next action, number multi-step work, suppress tangents, use concrete time estimates and remove ceremonial openings or closers. Its own documentation says no diagnosis is needed. The skill is an output style for reducing the distance between reading and acting, not a medical tool.

This distinction matters for search and procurement. It does not clone a brand voice or improve the factual basis of an article. It makes operational answers easier to scan and execute. That can be valuable in support, runbooks and internal delivery, but it can be too rigid for essays, sensitive messages or explanations that need nuance. Apply it by task, not as an unquestioned rule for every channel.

3. book-to-skill: turn reference material into on-demand knowledge

book-to-skill accepts a book, document folder or collection of sources and generates a skill with a core index plus chapter, glossary, pattern and cheat-sheet files. The agent can load the relevant section instead of receiving the whole source on every request. The project explicitly lists brand and design systems as one use case.

This solves a knowledge-packaging problem, not an editorial-quality problem. A generated skill inherits omissions, contradictions and licensing constraints from its inputs. Review the output against the original documents, keep provenance, and do not turn material you cannot lawfully reuse into a shared company asset. For a brand system, start from approved examples, not every text the company has ever published.

How should the three layers work together?

1. Ground the task. Supply the product facts, approved claims, audience, channel and desired reader action.
2. Load only relevant knowledge. Use a small reference or a progressively disclosed skill for the applicable product, market and voice rules.
3. Shape the response. Apply concise, action-first structure when the task needs it. Do not impose support-answer formatting on a long-form article.
4. Edit the draft. Run a focused slop and clarity pass after the argument exists, with an explicit rule to preserve facts, qualifications and author voice.
5. Verify and approve. Check every consequential claim against its source, score the output and keep a person responsible for publication.

The sequence matters. Style cleanup before grounding can make unsupported copy sound more credible. Grounding without editorial review can leave correct information unreadable. One skill can call or compose with another, but each should own a coherent responsibility.

How do you build a company writing skill?

The open Agent Skills specification defines a skill as a directory with a required SKILL.md file and optional scripts, references and assets. Its progressive-disclosure model loads basic metadata first, full instructions on activation and supporting files only when needed. That structure is a good fit for a writing system because not every task needs every product sheet, legal rule or channel template.

1. Choose one bounded job. Start with “edit an evidence-backed B2B article to our house style,” not “write everything for marketing.”
2. Collect good and bad examples. Use approved source text, explain why each example passes or fails, and remove confidential customer material.
3. Separate invariants from preferences. Claims need evidence and legal wording must stay intact. Sentence rhythm and contractions can vary by channel.
4. Write the smallest useful procedure. The official skill creator guidance recommends coherent scope, moderate detail and progressive disclosure instead of an exhaustive master prompt.
5. Bundle deterministic checks. A script can catch banned phrases, missing citations, title length, broken links or unresolved placeholders more reliably than another prose instruction.

How do you evaluate whether the writing skill works?

Build a test set before tuning the skill. Include short support answers, technical explanations, product pages, executive summaries, weak source material and cases where the correct behavior is to ask for evidence. Run each case with and without the skill on the same model snapshot.

The official Agent Skills evaluation guide recommends learning from failed assertions, human feedback and execution transcripts. That is the right loop for writing too. If the agent ignores a rule, clarify it. If it follows the rule and quality falls, remove or narrow the rule. A larger instruction file is not a success metric.

What are the production risks?

• Instruction supply chain. A mutable repository can change after review. Pin a commit or release, inspect scripts and retest upgrades.
• False confidence. Natural prose can make a weak claim look trustworthy. Keep citation and approval gates outside the style pass.
• Voice flattening. A universal list of forbidden patterns can erase intentional rhythm, humor or cultural differences. Localize judgment, not only words.
• Source leakage. Brand books, customer examples and internal documents may contain confidential or personal data. Control what enters model context.
• Rule collisions. Multiple active skills can disagree about length, tone or formatting. Define precedence and test the combined stack.

A production workflow needs version control, a model and prompt change log, source permissions, evals, approval roles and rollback. Our AI agent SLA template shows how to make those operating boundaries measurable.

Install a writing skill or build a custom system?

The commercial decision is not “prompt versus agency.” It is whether the cost of repeated manual correction now exceeds the cost of owning a tested content system. Wavect's AI enablement team can turn an approved editorial process into a versioned agent workflow. The Twinsoft AI case study shows the delivery discipline behind a production AI product, while our prototype-to-production guide helps scope the engineering gap. For a concrete architecture and pilot estimate, book an AI workflow review.