CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

obliteratus

OBLITERATUS: abliterate LLM refusals (diff-in-means).

Quality

73%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Low

Low-risk findings worth noting

Fix and improve this skill with Tessl

tessl review fix ./optional-skills/mlops/obliteratus/SKILL.md
SKILL.md
Quality
Evals
Security

Quality

Content

85%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

The body is a strong operational document: fully executable commands, explicit validation with feedback loops, and dense tables that assume Claude's competence. Its main weaknesses are discoverability of the bundle — an orphaned methods-guide.md reference file and nonexistent template paths — plus minor marketing-style padding.

Suggestions

Reference 'references/methods-guide.md' from the method-selection section (e.g., 'See references/methods-guide.md for theory and per-method details') so the existing bundle file is discoverable and the inline method list can be slimmed.

Remove or fix the 'templates/*.yaml' references — no templates directory exists in the bundle, so these paths are broken pointers.

Trim the 'What's inside' stats line and consolidate the 9-methods list with the method-selection table to reduce duplication.

DimensionReasoningScore

Conciseness

The body is dense and operational — commands, flag tables with defaults, VRAM tiers, pitfalls — with no explanations of concepts Claude already knows. Minor trimmable padding exists ('What's inside' stats line, Video Guide section, and the 9-methods list duplicating the method-selection table), which keeps it below the 'every token earns its place' anchor 5 but above anchor 3's 'several unnecessary explanations'.

4 / 5

Actionability

Fully executable throughout: 'obliteratus obliterate <model_name> --method advanced --quantization 4bit --output-dir ./abliterated-models', a complete hardware-check Python snippet, a copy-paste transformers test, a flag table with defaults, and concrete verification thresholds (refusal rate < 5%, perplexity change < 10%). Covers the common cases exactly as the anchor-5 example does.

5 / 5

Workflow Clarity

Steps 1–7 form a clear install → hardware → choose → run → verify → use sequence, with Step 6 as an explicit validation checkpoint and error-recovery feedback loops ('If refusals persist' and 'If coherence is damaged' each list specific parameter changes to try). This matches the anchor-5 pattern of explicit validation plus feedback loops; not 4 because checkpoints are comprehensive rather than minor-gapped.

5 / 5

Progressive Disclosure

Section structure is good and 'references/analysis-modules.md' is clearly signaled one level deep, but the bundle's 'references/methods-guide.md' is never referenced in the body, three cited 'templates/*.yaml' paths do not exist in the bundle, and method details that live in the reference are partially duplicated inline. These are more than 'minor organization gaps' (anchor 4) but the structure itself is sound, fitting anchor 3.

3 / 5

Total

17

/

20

Passed

Description

52%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

The description is concise and highly distinctive, with a clear statement of what the skill does, but it is minimal: only one action is named, common trigger synonyms ('uncensor', 'abliteration', 'refusal removal') are missing, and there is no 'when to use' guidance at all. Adding an explicit trigger clause would substantially raise completeness and trigger-term quality.

Suggestions

Add an explicit 'Use when...' clause, e.g. 'Use when the user wants to uncensor/abliterate a model, remove refusal or guardrail behaviors, or create an uncensored version of Llama, Qwen, or Mistral.'

Include natural trigger synonyms users actually say — 'uncensor', 'abliteration', 'refusal removal', 'guardrails' — to improve keyword coverage.

Name one or two more concrete capabilities (e.g., weight projection without retraining, verification metrics) to lift specificity beyond a single action.

DimensionReasoningScore

Specificity

The description names the domain ('LLM refusals') and a single concrete action ('abliterate') with one technique ('diff-in-means'), but coverage is minimal — no weight projection, direction extraction variants, verification, or output model handling. This matches anchor 2 ('actions are minimal or generic') better than anchor 3, whose example lists two concrete actions.

2 / 5

Completeness

A clear 'what' is present ('abliterate LLM refusals (diff-in-means)') but there is no 'Use when...' clause or equivalent trigger guidance, capping completeness at 3 per the judging guidelines. It is not 2 because the 'what' is concrete rather than vague.

3 / 5

Trigger Term Quality

'abliterate' and 'refusals' are natural terms users would say, but common variations are absent — 'uncensor', 'abliteration', 'refusal removal', 'guardrails', and specific model names do not appear. Fits anchor 3 ('some relevant keywords but missing common variations or synonyms') rather than 4, which requires only a few missing natural terms.

3 / 5

Distinctiveness Conflict Risk

'Abliterate LLM refusals' carves a clear, niche trigger space that no plausible sibling skill (serving, conversion, tokenizers) would claim. Not 4, since overlap risk with closely related skills is essentially nil given the highly specific terminology.

5 / 5

Total

13

/

20

Passed

Validation

81%

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 13 / 16 Passed

Validation for skill structure

CriteriaDescriptionResult

metadata_version

'metadata.version' is missing

Warning

metadata_field

'metadata' should map string keys to string values

Warning

frontmatter_unknown_keys

Unknown frontmatter key(s) found; consider removing or moving to metadata

Warning

Total

13

/

16

Passed

Repository

NousResearch/hermes-agent

Path

optional-skills/mlops/obliteratus/SKILL.md

Reviewed

Table of Contents

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.