voice-layer

voice-layer

A local-first personal voice layer for AI agents.

Bring your voice to any agent. Adapt to any audience without losing yourself.

Humanize AI-generated drafts by rewriting them in your own calibrated voice: PR comments, Slack replies, docs, RFCs, and emails.

voice-layer gives Claude Code and Codex a reusable local profile for writing like you: direct where you are direct, careful where you are careful, and shaped for the channel without flattening into generic AI prose. It is a local-first AI writing assistant for personal voice, not a generic humanizer prompt.

generic AI draft + local voice-profile.md -> PR comment | Slack reply | RFC | email

The demo uses only the synthetic profile in this repo.

Install

Homebrew install:

brew tap ymeiri/voice-layer
brew install voice-layer
voice-layer install --agent both
voice-layer doctor --agent both

Remove skills while preserving your profile:

voice-layer uninstall --agent both

Delete local profile data only when you explicitly mean it:

voice-layer purge --yes

Clone-based install also works:

./voice-layer install --agent both
./voice-layer doctor --agent both

This repo currently ships the reference implementation as two Agent Skills:

• write-in-my-voice: drafts or rewrites Slack replies, emails, PR descriptions, review comments, docs, design docs, RFCs, issue updates, and similar text.
• calibrate-my-voice: builds a local voice profile from samples you explicitly approve.

The repo and plugin packages use the voice-layer name. The commands are verbs: one for writing, one for calibration.

Current packaged support is Claude Code and Codex. "Bring your voice to any agent" is the long-term direction; Cursor and other agents are not claimed as supported until their package and invocation behavior are verified. See docs/support-matrix.md.

Why this is not just a prompt:

• Local profile: inspectable voice-profile.md, stored outside the repo.
• Consent-first calibration: the agent asks before reading private sources.
• Channel-aware output: PR comments, Slack replies, docs, RFCs, and email get different shapes from the same facts.
• Anti-slop cleanup: repeated model phrasing, filler, and punctuation tells are explicitly checked.
• Not detector bypass: voice-layer is for authentic personal communication, not evading AI detectors or hiding provenance.
• CI privacy guard: executable repo code is scanned for silent network/telemetry paths.

Why This Exists

AI agents are good at producing text. They are much worse at preserving the person behind the text.

Most AI-polishing tools remove a few cliches. That is not enough. Real writing depends on voice, channel, audience, context, document shape, and boundaries. A message can be more corporate-friendly, more direct, or easier for a different audience to read without erasing the person who wrote it. A design doc can also keep the user's structure, depth, risk framing, and decision style.

If you are searching for an AI text humanizer, voice-layer is the stricter version of that idea: it humanizes AI-written text by applying an explicit, auditable voice profile that you control.

voice-layer treats style as a portable profile:

voice + vibe + channel + audience + guardrails -> usable draft

What Is Different

• Portable profile: your calibrated voice is stored as an inspectable local file that agents can reuse.
• Consent-first calibration: the agent must ask before reading private sources such as Slack, email, docs, PRs, or project tools.
• Audience adaptation: the goal is to adapt for readers without turning cultural or regional language into a costume.
• Documentation style: design docs, RFCs, ADRs, runbooks, and wiki pages can preserve how the user structures context, alternatives, risks, and decisions.
• Anti-slop cleanup: the repo includes a public catalog of patterns that make AI writing sound fake: AI_TELLS.md.
• Spec-first direction: the profile model is documented in VOICE_PROFILE_SPEC.md.

Zero-Data Demo

You can test the core loop without connecting Slack, Google Docs, GitHub, email, or any private workspace. The repo includes a synthetic profile at tests/fixtures/profiles/valid-profile.md.

Install the Claude Code skill from this checkout, then launch Claude Code with the bundled profile:

./voice-layer install --agent claude --mode symlink
VOICE_LAYER_PROFILE="$PWD/tests/fixtures/profiles/valid-profile.md" claude

Then ask Claude Code:

/write-in-my-voice draft a PR comment from this:

Change parseWidgetConfig() in src/widgets/config.ts.
Goal: fix FOO-123.
Behavior is behind ENABLE_WIDGET_V2=true.
Tests: npm test -- widgets/config.test.ts.
Concern: we should make sure invalid widget types still return the existing validation error.

A real zero-data Claude Code run produced:

Patch parseWidgetConfig() in src/widgets/config.ts for FOO-123. Gated on
ENABLE_WIDGET_V2=true.

One thing to confirm before this lands: invalid widget types should still
return the existing validation error, not fall through to the new path. Worth
an explicit test case alongside npm test -- widgets/config.test.ts.

The same facts can be shaped for another channel:

/write-in-my-voice make this suitable for a short Slack update to the team:

Change parseWidgetConfig() in src/widgets/config.ts.
Goal: fix FOO-123.
Behavior is behind ENABLE_WIDGET_V2=true.
Tests: npm test -- widgets/config.test.ts.
Concern: invalid widget types should still return the existing validation error.

Output:

Patching parseWidgetConfig() in src/widgets/config.ts for FOO-123, gated on
ENABLE_WIDGET_V2=true. Tests: npm test -- widgets/config.test.ts. Watch item:
invalid widget types should still hit the existing validation error. Flagging
in case anyone sees otherwise.

The broader synthetic example set lives in examples/README.md. No examples are based on a real person.

Clone Install

From a clone of this repo, if you are not using Homebrew:

./voice-layer install --agent both
./voice-layer doctor --agent both

Restart Claude Code or Codex after installing.

The CLI links the skills into:

• Claude Code: ~/.claude/skills
• Codex: ~/.agents/skills

It also creates the local profile path:

~/.config/voice-layer/voice-profile.md

Preview file changes before installing:

./voice-layer install --agent both --dry-run

Remove skill links without deleting your profile:

./voice-layer uninstall --agent both

Delete local profile data only with an explicit purge:

./voice-layer purge --yes

For Homebrew, plugin/marketplace installs, uninstall, and troubleshooting, see docs/install.md.

First Use

Before calibration, write-in-my-voice can still remove generic AI writing patterns, but it does not know your voice yet.

Claude Code:

/write-in-my-voice rewrite this Slack reply:
Great question! It is important to note that we could potentially leverage this robust approach.

Codex:

Use $write-in-my-voice to rewrite this Slack reply:
Great question! It is important to note that we could potentially leverage this robust approach.

Calibrate

Start calibration when you want the skill to learn your style:

/calibrate-my-voice build my local voice profile.

The calibrating agent should first inspect what source tools are available in the current session, then ask which sources it may use. It can work from pasted samples, exports, local git commits, code reviews, chat, email, docs, or project-management tools such as Linear, Jira, Asana, Monday.com, or internal trackers.

It must not read any source until you approve the source, scope, date range, author filter, and retention mode.

Voice, Vibe, Audience

voice-layer separates stable identity from temporary adaptation:

• Voice: the user's calibrated writing baseline.
• Vibe: a temporary overlay such as corporate-friendly or chill-rock-star.
• Channel: Slack, email, PR review, docs, issue tracker, or another surface.
• Audience: the reader and how much context or formality they need.
• Document shape: the expected structure and depth for docs, RFCs, ADRs, runbooks, and design docs.
• Localization: clarity and convention adjustments, not cultural mimicry.

The key rule:

Voice is portable. Culture is not a costume.

Privacy

This repo does not include a hosted service, telemetry, or personal samples. CI runs a static no-silent-telemetry guard over executable repo code:

voice-layer is not an AI detector bypass tool. It is meant to help users communicate honestly in their own style, with local profile data they can inspect and delete.

python3 scripts/check_no_silent_telemetry.py

The generated profile stays local by default at:

~/.config/voice-layer/voice-profile.md

Read PRIVACY.md before calibrating from private workspaces.

Validate

python3 scripts/sync_core.py --check
python3 scripts/validate_repo.py
python3 scripts/check_no_silent_telemetry.py
python3 scripts/validate_profile.py ~/.config/voice-layer/voice-profile.md
python3 scripts/evaluate_examples.py
python3 scripts/evaluate_behavior.py
python3 scripts/evaluate_behavior.py --print-pilot-prompts
python3 -m unittest discover -s tests

Project Docs

• AI_TELLS.md: public catalog of common AI writing patterns.
• VOICE_PROFILE_SPEC.md: portable profile model.
• core/profile/schema.json: machine-readable profile schema.
• docs/install.md: install, plugin, and uninstall details.
• docs/evaluation.md: current evaluation tiers.
• docs/visual-identity.md: logo, visual language, and Nano Banana prompt pack.
• docs/launch-playbook.md: launch channels, copy, and public-posting guardrails.
• evals/behavior/scenarios.json: behavioral eval scenarios for writing, calibration, privacy, and audience adaptation.
• scripts/validate_profile.py: local voice profile validation.
• docs/support-matrix.md: verified support status.
• docs/implementation-roadmap.md: path to a complete multi-agent implementation.

References

• Agent Skills specification
• Codex skills
• Claude Code plugins