--- license: other license_name: polyform-noncommercial-1.0.0 license_link: https://polyformproject.org/licenses/noncommercial/1.0.0/ pretty_name: "speak-human - make AI writing readable without flattening the facts" language: - en - zh tags: - agent-skill - agent-skills - claude-skill - claude-code - skill - agentskills - prompt-engineering task_categories: - text-generation size_categories: - n<1K viewer: false --- > **speak-human** - Make public copy clearer and more human without flattening facts, context, intent or technical meaning. > > **Mirrored in three places, same version everywhere:** > [Hugging Face](https://huggingface.co/datasets/LucioLiu/speak-human) (you are here) · [GitLab](https://gitlab.com/LucioLiu/speak-human) · [GitHub](https://github.com/LucioLiu/speak-human) > > **Get the files** - GitLab is the most reliable plain-git route: > > ```bash > git clone https://gitlab.com/LucioLiu/speak-human.git > > # or from this page > hf download LucioLiu/speak-human --repo-type dataset --local-dir ./speak-human > ``` > > Then drop the folder into your agent's skills directory - `~/.claude/skills/speak-human/` for Claude Code, `~/.agents/skills/speak-human/` for Codex, or your client's equivalent. > > *Note: links to github.com inside the text below point to the GitHub mirror; if one does not resolve, use the GitLab mirror above - it carries the same content.* > > **Licence: PolyForm Noncommercial 1.0.0** - the `LICENSE` file in this repo is authoritative. --- **English** | [简体中文](README.zh-CN.md) ![Speak Human — one speaker, one reader, one purpose, one clear message](docs/speak-human-card.svg) # Speak Human Speak Human teaches an agent to turn complex, compressed, awkward, or bureaucratic writing into something a specific person can understand and act on. It works on status updates, technical explanations, decisions, reports, emails, README files, career material, replies, and public posts. It does not make everything casual. It keeps the facts, intent, technical depth, emotion, and formality that still serve the reader. > One speaker. One reader. One purpose. One clear message. [![Tests](https://github.com/LucioLiu/speak-human/actions/workflows/test.yml/badge.svg)](https://github.com/LucioLiu/speak-human/actions/workflows/test.yml) ## How it works This repository is an Agent Skill, not a standalone writing application. 1. A compatible host sees a human-facing writing task and loads [`SKILL.md`](skills/speak-human/SKILL.md). 2. The agent identifies the speaker, reader, purpose, desired outcome, exposure level, and facts that cannot change. 3. It finds the main message and restores only the context the reader needs. 4. It translates compressed or technical language through: `observation → meaning → practical effect → next action or condition` 5. It rewrites the text in the right voice and form, then checks facts, intent, depth, privacy, and edit scope. The optional Python helper only checks a few exposure-sensitive leaks. It does not write the copy. ## The core is not social-media operations Social media is one optional adapter. The general method comes first. | Job | What Speak Human makes clear | |---|---| | Status update | What is done, what proves it, what remains, and what happens next | | Technical explanation | Mechanism, example, use, prerequisites, failure behavior, and scope | | Decision or recommendation | The recommendation, evidence, tradeoff, and revisit condition | | Report, handoff, or email | Current state, request, constraints, owner, deadline, and next step | | README or release note | What changed, who is affected, how to start, tested scope, and limits | | Resume or portfolio | Problem, action, evidence, result, and honest validation scope | | Conversation or reply | The person's actual point, a direct answer, and the right relationship tone | | Social post | Original context, reader value, translation, emotion, evidence, and format | ## A concrete example Before: > The synchronization has been completed, the hashes of the three copies are consistent, and runtime discovery is pending restart. After: > The same skill is now in the project repository, the team library, and the local runtime. All three copies contain the same files with matching hashes, so no version drift was introduced. The runtime will recognize the new version after it restarts. The second version does not merely swap words. It names the objects, explains what the evidence proves, and tells the reader what remains. More worked examples are in [`references/core-method.md`](skills/speak-human/references/core-method.md). ## Install ### Agent Skills CLI ```bash npx skills add LucioLiu/speak-human --skill speak-human -g -a codex --copy -y ``` ### Codex skill installer Ask Codex: ```text $skill-installer install https://github.com/LucioLiu/speak-human/tree/main/skills/speak-human ``` Restart Codex after installation so it discovers the skill. ### Manual copy Clone the repository, then copy [`skills/speak-human`](skills/speak-human) into the host's user-level skills directory under the same `speak-human` name. ## Use it ```text Use $speak-human to rewrite this progress report for a smart reader outside my field. Lead with the current result, explain what the test evidence actually proves, and keep the remaining limitation. ``` ```text Use $speak-human to explain idempotency to a product manager. Keep the technical term, give one concrete example, and say when a retry is safe. ``` ```text Use $speak-human to edit this email. Put the request and deadline first, but keep it professional rather than chatty. ``` For a normal request, the skill returns usable wording first. It exposes rule IDs or an audit trail only when asked. ## Optional leak checker ```bash python skills/speak-human/scripts/check.py \ --exposure external \ --text "The draft is stored at C:\Users\someone\private\draft.md" ``` The checker treats exposure levels differently: a local path may be useful in an internal handoff and unsafe in a public document. It heuristically flags common credential-shaped assignments and bearer tokens at every exposure level. This small helper is not a complete secret scanner. ## Package layout ```text skills/speak-human/ ├── LICENSE.txt ├── SKILL.md ├── SKILL.zh-CN.md ├── agents/openai.yaml ├── assets/review-output-template.md ├── references/ │ ├── communication-modes.md │ ├── core-method.md │ ├── core-rules.json │ ├── language-signals.md │ ├── social-media.md │ └── voice-profile.schema.json └── scripts/check.py ``` Private author profiles are not distributed. A caller may supply a separate profile that follows the public schema. ## Version and verification Current version: **2.1.1**. Version 2.1 keeps constraints backstage unless the reader needs them, stops social posts from being forced into a lesson or polished ending, and preserves supported sharpness and unfinished edges. Tests cover package structure, bilingual contract markers, general trigger boundaries, fixed exposure-aware leak cases, version markers, and known private-material exclusions. Version 2.1.1 is a leak-checker maintenance patch; it does not change the writing method or the evaluation scores below. These tests verify the published skill contract. Output quality is evaluated separately with answer-isolated authorship tests; see [the evaluation method and current evidence](docs/evaluation.md). ## Privacy The helper has no network dependency or telemetry. See [PRIVACY.md](PRIVACY.md) before using a private voice profile. ## License Source available under the [PolyForm Noncommercial License 1.0.0](LICENSE). Personal and non-commercial use is permitted under its terms. This is not an OSI open-source license. --- > Both language editions are maintained together. If they diverge, the English `SKILL.md` and machine-readable core rules define the executable package contract.