| name | personal-humanizer-maker |
| description | Feed in a Korean or English sample of your own writing, and this meta-skill (a skill factory) decomposes its style and argument flow with a quantitative profiler plus one specialist agent per axis, then auto-generates a personal humanizer skill. The output is a standalone humanize-NAME skill for the current host: Claude Code emits under ~/.claude/skills, while Codex emits under the Codex skill root. It includes SKILL.md, style_metrics.py with the author's baselines baked in, and before/after examples mined from the author's own text. A meaning-invariance covenant is injected into every generated skill. Triggers — "make a personal humanizer from my writing", "build a humanizer skill in my voice", "personal humanizer maker", "profile my writing style into a skill", "내 문체로 다듬는 스킬 만들어줘", "이 글로 개인 휴머나이저 빌드", "내 voice 스킬 제작", "샘플 넣으면 내 문체 뽑아주는 거". Rewriting a single text is the job of an already-generated personal skill; THIS skill is the factory that stamps those personal skills out. |
personal-humanizer-maker — a personal-voice humanizer skill factory
Sample document → quantitative profile + qualitative analysis → a humanizer skill
(humanize-NAME/) that is that person's alone, emitted automatically. It is the
pipeline that stamps out — from any author's writing — the kind of hand-tuned instance a
human would otherwise build by hand (e.g. personal_humanize).
0. The covenant — meaning is invariant (injected into every output)
Every generated skill inherits the covenant in references/ironclad.md as its §0.
Whatever the author's style, this contract is fixed, author-independent.
- Facts, numbers, units, years, names, proper nouns, causality, and order never change — not by a single token.
- Citations, links, and footnotes are preserved verbatim. No reformatting, moving, or deleting.
- No new claims, evidence, or citations. Invent nothing that wasn't there.
- No over-editing. Only phrasing, rhythm, sentence joining, and within-paragraph arrangement change.
Three layers (code · reference · LLM)
| Layer | Does | Artifacts |
|---|
| CODE (deterministic, stdlib-only) | quantitative profiler · skill emitter · round-trip checker | scripts/profile_corpus.py, scripts/emit_skill.py, scripts/roundtrip_check.py, scripts/build_profile.py |
| REFERENCE (static knowledge) | style-dimension taxonomy (ko/en) · signal→axis map · covenant · templates | references/taxonomy.{ko,en}.md, references/signal-map.md, references/ironclad.md, templates/* |
| LLM / multi-agent (interpretive) | axis specialists · synthesizer · fidelity auditor | orchestration below |
Pipeline
sample doc + language (ko/en)
│ [CODE] profile_corpus.py
▼
quant_profile.json (7-axis distributions: sentence length · ending mix · connective density · passive rate · gloss rate · formatting …)
│ [MULTI-AGENT] one specialist agent per axis, fanned out ← references/taxonomy.{lang}.md
▼
dimension_profiles[] (per-axis value + confidence + rules + exemplars mined from the author)
│ [MULTI-AGENT] synthesizer (barrier)
▼
style_profile.json (unified rules + calibrated baselines + canonical examples) + style_profile.md
│ [CODE] emit_skill.py + templates ← references/ironclad.md (covenant injected)
▼
humanize-NAME/ skill package
│ [CODE+AGENT] roundtrip_check.py + fidelity audit ← automatic safety gate
▼
PASS → ship / FAIL → widen bands · demote low-confidence axes · re-emit once → ship with CONFIDENCE note
Language selection (ko/en)
At the start, pick Korean or English. That choice selects taxonomy.{lang}.md, the
profiler's language module, and the emitted skill's language mode together. The shared
skeleton + per-language module design means a third language is one taxonomy file plus one
profiler module.
Automatic-mode safety (no human review)
With no reviewer in the loop, overfitting is defended in code.
- Thin-corpus check: below the character threshold, confidence is demoted and bands widen.
- Per-axis confidence: based on exemplar count. Low-confidence axes ship as
advisory, not strict.
- Round-trip gate: right after emit, the skill is applied to a neutralized held-out text to confirm (a) the baseline bands converge and (b) meaning is preserved. On failure it re-emits once; if it still fails, it ships with a
CONFIDENCE.md warning. An overfit skill is never shipped silently.
Orchestration procedure (how to run)
When invoked, proceed in order. CODE steps are deterministic; agent steps are interpretive.
0. Gather input. Confirm the language (ko/en), and take the sample path(s) + a
profile_name (slug) + a display_name. Warn if the corpus is thin (below threshold).
1. Quantitative profile (CODE).
python3 scripts/profile_corpus.py --lang <ko|en> <sample...> -o runs/<name>/quant_profile.json
2. Axis-specialist fan-out (multi-agent). Launch one subagent in parallel per axis
(sentence_architecture, register_modality, lexical_register, cohesion_argument, stance_voice, figuration, formatting). Prompt each with:
You are the specialist. Read: (a) the section of references/taxonomy.<lang>.md, (b) the author sample, (c) the slice of quant_profile.json. Emit one dimension_profile JSON (schema schemas/dimension_profile.schema.json) describing how this author handles the axis. Include — observations (with evidence), confidence (by exemplar count: ≥8 high / 3–7 medium / <3 low), rules for a rewriter (imperative, with the author's observed values filled in), and before/after exemplars copied verbatim from the sample (after = a real author sentence, before = a plainer same-meaning paraphrase). Do not invent axes — fill the taxonomy frame. Facts/content are not your concern — only the shape of the style. Do not manufacture a rule for a trait the author does not exhibit.
Save each agent's output to runs/<name>/dims/<axis>.json. (Use parallel(7 agents) when a Workflow is available, otherwise 7 concurrent Agent-tool calls.)
3. Assemble (CODE). Baseline bands are derived by code, not the LLM, from the quant profile (reproducibility).
python3 scripts/build_profile.py --quant runs/<name>/quant_profile.json --dims runs/<name>/dims \
--name <name> --display "<display_name>" -o runs/<name>/style_profile.json
4. Emit (CODE).
python3 scripts/emit_skill.py runs/<name>/style_profile.json --host auto
--host auto emits to the active host's skill root: ~/.claude/skills for Claude Code,
${CODEX_HOME:-~/.codex}/skills for Codex. Use --host claude, --host codex, or -o
to override explicitly.
5. Round-trip gate (CODE + agent). Apply the fresh skill to a neutralized held-out
text to confirm baseline convergence + meaning preservation. On failure, widen bands /
demote low-confidence axes / re-emit once; if it still fails, ship with CONFIDENCE.md.
Optional synthesizer agent. When axis outputs conflict (e.g. the lexical and
figuration axes propose opposing rules) or exemplars are excessive, a synthesizer agent
may tidy the dims before build_profile.py. The default code-merge path is enough.
Components
| Component | File | Role |
|---|
| Contract schemas | schemas/*.schema.json | code↔agent glue; installed copies are bundled inside the skill |
| Quantitative profiler | scripts/profile_corpus.py | 7-axis distributions → quant_profile.json |
| Taxonomy references | references/taxonomy.{ko,en}.md + signal-map.md | per-axis rule frame + band derivation |
| Profile assembler | scripts/build_profile.py | derive baselines + merge dims → style_profile.json |
| Skill emitter | scripts/emit_skill.py + templates/* | style_profile.json → humanize-<name>/ |
| Round-trip gate | scripts/roundtrip_check.py | convergence check + relaxation |
Contracts (schemas/)
The three contracts that join code and agents:
quant_profile.schema.json — the profiler's output.dimension_profile.schema.json — one axis specialist's output.style_profile.schema.json — the synthesizer's output, the emitter's sole input.
