a portable skill for Claude Code, opencode & any agent · v3.0.0
It teaches a language model how to talk like a person. Not how to sound like one — that part's easy, and it's what most failures already do. This works on the shape underneath: when to be short, when to sit with something, when to push back, when the right reply is just "oh".
the difference between sounding human and being shaped like one.
Watch the shape of a real reply.
Same skill, different moments. It matches the register, leads with the human, and knows when a short answer is the honest one.
Two ways to be human.
A load-time router picks one; ambiguity defaults to A, the warm default. Both share one AI-tell taxonomy and one self-audit pass.
You are the friend, replying in real time — matching register, leading with acknowledgment, honoring silence, pushing back when a real friend would.
Paste AI-drafted prose and get it back human, in your voice — the use case the skill used to decline. Strips the tells without hollowing it out.
A self-audit on every reply.
A big ruleset with no checking step leaves residue. Before every finalized reply, in both modes, the skill re-reads its own draft:
Read the draft as a skeptic. What reads as AI? Enumerate against the tell taxonomy.
Cut the specific tells — targeted, not a blanket reword. Never invent facts to fill a gap.
Still a pulse? A tell-clean but sterile draft fails. Stripping is necessary; soul is the point.
The v3.0 scorecard.
A blind, independent oracle scored a stratified sample of replies produced under v3.0 — a mix of the new cases and a Mode-A regression set. Real numbers, mixed edges and all.
One skill, every agent.
iamhumans is a single portable SKILL.md — no runtime, no dependency. Drop it wherever your agent reads skills, or just paste it into a system prompt.
git clone https://github.com/hoainho/iamhumans
# install as a Claude Code skill
mkdir -p ~/.claude/skills/iamhumans
cp -R iamhumans/SKILL.md iamhumans/references ~/.claude/skills/iamhumans/
# then just talk — Claude Code auto-loads it on human-shaped turns,
# or hand it a draft: "make this sound less like a bot"git clone https://github.com/hoainho/iamhumans
cd iamhumans
# install as a local opencode skill (symlink stays in sync)
mkdir -p ~/.opencode/skills/iamhumans
ln -s "$PWD/SKILL.md" ~/.opencode/skills/iamhumans/SKILL.md# Cursor, Windsurf, a raw API loop, your own harness —
# iamhumans is prose, so any model can use it:
# 1. open SKILL.md
# 2. paste it into the system prompt / instructions
# 3. load it when the conversation is human-shaped
curl -O https://raw.githubusercontent.com/hoainho/iamhumans/main/SKILL.mdMCP roadmap — a thin MCP server that serves the skill to any MCP-capable provider is on the roadmap. Today the skill is the product; the MCP wrapper is a distribution convenience, not a dependency.
v3.0 is measured on a blind-graded 30-case sample — aggregate 82.7/100, 27/30 PASS, 0 hard-fails (the scorecard above), plus schema dry-run, lint, and a Mode-A regression pass in which the one miss was a door-reopener that has since been fixed. It has not yet been re-scored on the full 429-case oracle harness, and the sample keeps its mixed edges — two de-AI replies and one register pivot scored below the bar, and none of it is hidden.
The same model lineage authored the skill, the cases, the replies, and served as oracle judge — a contamination named from the start. Book notes are distilled from training-time exposure, not real-time reading; every claim is marked as paraphrase, with no fabricated page numbers.