Heretic Abliteration
Heretic — fully automatic LLM censorship removal (abliteration) and serving the result via Ollama. Wraps p-e-w/heretic: directional ablation (difference-of-means refusal directions, orthogonalizing attn.o_proj + mlp.down_proj) with an Optuna/TPE search that co-minimizes refusal count and KL divergence. Automates the full local workflow with the dependency + device fixes that make it actually run: clean-venv install (+ kernels/jinja2 fixes), device-correct run (CPU on Apple Silicon — MPS is blocked by a PyTorch GQA matmul bug; CUDA for real models), and the HF→GGUF→Ollama bridge. Includes a no-compute fast path: pull a pre-abliterated GGUF straight into Ollama. Use when: (1) decensoring / uncensoring an open-weight LLM locally, (2) running heretic-llm and hitting install or MPS/CUDA errors, (3) converting a Heretic/abliterated HF model to GGUF and serving it via Ollama CLI, (4) pulling a ready-made abliterated model into Ollama, (5) understanding abliteration as an evaluator-governed (EGRI) optimization loop. Triggers on: 'heretic', 'abliteration', 'abliterate', 'decensor', 'uncensor', 'remove refusals', 'refusal direction', 'heretic-llm', 'uncensored model', 'ollama abliterated', 'run heretic locally'. NOT FOR training/fine-tuning safety in, or for API-only models (abliteration needs local weights).
Security AStatic scan found no risk patternsHow grading works ›
From the source SKILL.md
Remove refusals from open-weight LLMs with [p-e-w/heretic](https://github.com/p-e-w/heretic), then serve the result locally through the Ollama CLI. Heretic combines directional ablation ("abliteration") with an Optuna/TPE search that co-minimizes refusals and KL divergence — decensoring the model while preserving its intelligence, no fine-tuning required.
What this skill does
Heretic Abliteration is a community-contributed Claude Code skill in the product-misc sub-category. It ships as a SKILL.md file that Claude Code auto-discovers under ~/.claude/skills/heretic-abliteration/ and loads when your prompt matches the skill's trigger.
When to invoke it: Use when: (1) decensoring / uncensoring an open-weight LLM locally, (2) running heretic-llm and hitting install or MPS/CUDA errors, (3) converting a Heretic/abliterated HF model to GGUF and serving it via Ollama CLI, (4) pulling a ready-made abliterated model into Ollama, (5) understanding abliteration as an evaluator-governed (EGRI) optimization loop. Triggers on: 'heretic', 'abliteration', 'abliterate', 'decensor', 'uncensor', 'remove refusals', 'refusal direction', 'heretic-llm', 'uncensored model', 'ollama abliterated', 'run heretic locally'.
Who uses this skill
The Heretic Abliteration Claude Code skill is built for product managers, product designers, and cross-functional teams planning, shipping, and measuring product features. It's part of ClaudSkills (also referred to as Claude Skills or Claude Code Skills) — the open community-curated registry of 211,000+ SKILL.md files for Anthropic's Claude Code agent and the wider Claude ecosystem (Claude API, Claude Agent SDK).
How to install
Free
Manual install (2 steps)
mkdir -p ~/.claude/skills/heretic-abliteration
curl -L https://claudskills.com/skills/heretic-abliteration/SKILL.md \
-o ~/.claude/skills/heretic-abliteration/SKILL.md
Or just download SKILL.md directly and drop it into ~/.claude/skills/heretic-abliteration/. Claude Code auto-discovers it on next session.
Skills live at ~/.claude/skills/heretic-abliteration/SKILL.md on macOS/Linux, or %USERPROFILE%\.claude\skills\heretic-abliteration\SKILL.md on Windows. See the full install guide for step-by-step instructions.
Telegram
📱 Install from your phone or desktop Telegram
Open @claudskills_bot on Telegram, tap Open Desktop App, and the desktop app installs this skill for you. Or share the bot link with a colleague — they get the same one-tap install. Learn more →
Pro
One-click install via the desktop app
The ClaudSkills desktop app installs any skill directly into ~/.claude/skills/ with one click — no terminal required. Pro starts at $9/mo or $149 lifetime.
Pro
For the full experience including quality scoring and one-click install features for each skill — upgrade to Pro.
Frequently asked questions
How do I install the Heretic Abliteration Claude Code skill?
Install via the ClaudSkills desktop app (one click) or copy
SKILL.md from the source repository to
~/.claude/skills/heretic-abliteration/SKILL.md and restart Claude Code. Both flows are detailed at
claudskills.com/install/.
What does the Heretic Abliteration skill do?
Heretic — fully automatic LLM censorship removal (abliteration) and serving the result via Ollama. Wraps p-e-w/heretic: directional ablation (difference-of-means refusal directions, orthogonalizing attn.o_proj + mlp.down_proj) with an Optuna/TPE search that co-minimizes refusal count and KL divergence. Automates the full local workflow with the dependency + device fixes that make it actually run: clean-venv install (+ kernels/jinja2 fixes), device-correct run (CPU on Apple Silicon — MPS is blocked by a PyTorch GQA matmul bug; CUDA for real models), and the HF→GGUF→Ollama bridge. Includes a no-compute fast path: pull a pre-abliterated GGUF straight into Ollama. Use when: (1) decensoring / uncensoring an open-weight LLM locally, (2) running heretic-llm and hitting install or MPS/CUDA errors, (3) converting a Heretic/abliterated HF model to GGUF and serving it via Ollama CLI, (4) pulling a ready-made abliterated model into Ollama, (5) understanding abliteration as an evaluator-governed (EGRI) optimization loop. Triggers on: 'heretic', 'abliteration', 'abliterate', 'decensor', 'uncensor', 'remove refusals', 'refusal direction', 'heretic-llm', 'uncensored model', 'ollama abliterated', 'run heretic locally'. NOT FOR training/fine-tuning safety in, or for API-only models (abliteration needs local weights).
Is this skill free to install?
Yes. ClaudSkills is an open registry — every skill keeps its source repository's license, and manual install via copy is free. ClaudSkills Pro ($9/mo, $79/yr, or $149 one-time) adds one-click install via the desktop app and a multi-signal Quality Score.
When should I use the Heretic Abliteration skill?
Use Heretic Abliteration when your Claude Code task falls under the Product category — specifically in the product misc area. Claude Code auto-discovers installed skills and invokes the right one based on the task description, so you can also ask Claude directly (e.g. "use Heretic Abliteration" or describe the task and let Claude pick). Browse related skills at
/category/product/.
What is a Claude Code skill and how does the Heretic Abliteration skill fit in?
A Claude Code skill is a
SKILL.md file that lives under
~/.claude/skills/<name>/ and tells the Claude Code CLI agent how to perform a specific task (instructions, prompts, allowed tools). Skills are auto-discovered at session start. Heretic Abliteration is one of 67,000+ skills indexed in the open ClaudSkills catalog, classified under the Product category. Learn more at
/learn/what-is-a-claude-skill/.
Attribution & license
Cite this skill
If you reference this skill in a blog post, paper, or documentation, you can cite it as:
APA
broomva. (2026). Heretic Abliteration [Claude Code skill]. ClaudSkills. https://claudskills.com/skills/heretic-abliteration/
BibTeX
@misc{heretic-abliteration-2026,
author = {broomva},
title = {Heretic Abliteration [Claude Code skill]},
year = {2026},
publisher = {ClaudSkills},
url = {https://claudskills.com/skills/heretic-abliteration/}
}Embed this skill
Promote, attribute, or link this skill from your own README, blog post, or documentation. All three snippets are free to use — no sign-up, no API key. More distribution surfaces →
Badge
[](https://claudskills.com/skills/heretic-abliteration/?utm_source=badge&utm_medium=readme&utm_campaign=skill_badge)
<script>
<script src="https://claudskills.com/embed/heretic-abliteration.js" async></script>
<iframe>
<iframe src="https://claudskills.com/embed/heretic-abliteration.html" width="100%" height="160" frameborder="0" loading="lazy" title="ClaudSkills: Heretic Abliteration"></iframe>
Security scan
Grade A · scanned 2026-10-25 — free static scan against the OWASP Agentic Skills Top 10.
No risk patterns were found in any of the ten OWASP-aligned categories. How grading works ›
- ✓ Prompt injection
- ✓ Data exfiltration
- ✓ Supply chain
- ✓ Reverse shell
- ✓ Credentials
- ✓ Execution
- ✓ Filesystem
- ✓ Persistence
- ✓ Obfuscation
- ✓ Network
Show this grade on your repo (click to copy):
[](https://claudskills.com/skills/heretic-abliteration/#security)
More Product skills
Browse all Product skills in the ClaudSkills registry, or explore these other picks from the same category:
Part of Acreator Store — Adam Lankamer's AI tools:
PerfectStudio ·
Ucaption ·
UTagger ·
AutoXPoster ·
TestYourSkills ·
AutomationFlows ·
Au Naturel ·
Telegram @acreatorstore