human-likely
Make Claude write technical docs like a senior engineer writing to teammates — with trade-offs, war stories, and accountability. No AI fluff.
What is this?
A Claude Code Plugin that ships two complementary Skills for technical writing:
| Skill | When to use | What it does |
|---|
humanize-doc | Rewrite / polish / improve existing docs; complaints like "too AI-sounding", "talk like a human", "write like a senior engineer" | Polish existing text |
engineering-deepdive | Write a technical deep-dive, architecture analysis, design doc, engineering writeup, or code walkthrough from scratch | Generate new long-form technical content |
Recommended workflow: use engineering-deepdive to draft a question-driven, well-structured article with diagrams, then run humanize-doc to strip any remaining AI clichés and polish the tone.
Why does this exist?
LLMs default to a recognizable writing pattern: passive voice, feature lists, filler phrases ("综上所述", "旨在", "赋能", "全方位"), and zero decision rationale. The result reads correct but feels hollow — no one believes the author actually built or debugged the system.
human-likely fixes this by injecting two things into Claude's writing process:
- Decision traces — every non-trivial design choice shows the naive approach that failed and why the real approach was picked.
- Structural discipline — question-driven organization replaces feature lists; overview-detail-synthesis replaces "first... second... in conclusion...".
The difference isn't vocabulary. It's whether the reader finishes thinking "this person has actually shipped this" vs "this was generated".
Quick Start
Option 1: Install from this repo (recommended)
Register this repo as a single-plugin marketplace, then install:
# 1. Add marketplace (reads .claude-plugin/marketplace.json)
claude plugin marketplace add https://github.com/leslieJt/human-likely
# 2. Install the plugin
claude plugin install human-likely@human-likely
Launch claude in any project and start writing. Claude auto-activates the right Skill based on context, or trigger explicitly:
/human-likely:humanize-doc
/human-likely:engineering-deepdive
Option 2: Local development
Clone and mount directly with --plugin-dir (changes take effect immediately, no marketplace needed):
git clone https://github.com/leslieJt/human-likely.git
cd /your/project
claude --plugin-dir /path/to/human-likely
Option 3: Manual install as user-level Skill
Copy the skill directories into ~/.claude/skills/ — the Skills themselves don't depend on plugin context:
cp -r human-likely/skills/humanize-doc ~/.claude/skills/
cp -r human-likely/skills/engineering-deepdive ~/.claude/skills/
Skills in Detail
humanize-doc — Polish existing docs
Trigger phrases: "rewrite this", "polish", "too AI-sounding", "说人话", "像资深工程师那样写"
What it does:
- Diagnose — identify the audience, the real pain point, and where AI clichés hide.
- Extract technical skeleton — lock down facts that must not be lost (APIs, fields, numbers, constraints).
- Rewrite following four principles:
- Rigorous — exact numbers, specific function names, concrete paths. Never blur "2,847 QPS" into "fast".
- Human — inject decision rationale, trade-offs, and war stories. At least one "why we didn't pick the other approach" and one concrete gotcha.
- De-mechanized — delete all AI clichés ("综上所述", "旨在", "赋能", "全方位", "不仅...而且...").
- Balanced — professional but not stiff, conversational but not sloppy. 0-3 emoji max in a technical essay.
- Self-check against a strict checklist before delivery.
Example transformation:
| Before (AI-flavored) | After (human-likely) |
|---|
| 在当今快速发展的微服务架构领域，服务间的可靠通信至关重要。本方案旨在通过引入消息队列，全方位提升系统的稳定性与扩展性。 | 上个季度我们因为下游 HTTP 直连超时，连环挂了三次。这次把订单事件改走 Kafka，主要是为了把"下游慢"和"上游能不能下单"解耦——副作用是引入了消息重复的问题，所以下游必须做幂等。 |
engineering-deepdive — Generate from scratch
Trigger phrases: "deep-dive", "深度剖析", "architecture analysis", "design doc", "engineering writeup", "code walkthrough", "技术博客"
Five core writing patterns:
