constitutional-ai

Anthropic's method for training harmless AI through self-improvement. Two-phase approach - supervised learning with self-critique/revision, then RLAIF (RL from AI Feedback). Use for safety alignment, reducing harmful outputs without human labels. Powers Claude's safety system.

Quality

58%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Passed

No findings from the security scan

Fix and improve this skill with Tessl

tessl review fix ./backend/cli/skills/llm-tools/constitutional-ai/SKILL.md

Quality

Content

50%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 content is well-structured with concrete, illustrative code for both training phases and useful troubleshooting guidance, but it is held to mid-range scores by redundant concept explanations, code with undefined helpers, absent validation checkpoints for batch training, and no progressive disclosure into bundle files.

Suggestions

Add validation/verification checkpoints to the training workflows — e.g. evaluate the reward model's accuracy on a held-out preference set and measure harmlessness before proceeding to PPO — to lift workflow clarity above the batch-operation cap.

Make code examples fully executable by defining or stubbing the helper functions (create_dataset, parse_preferences, create_preference_dataset) and the undefined variables (model, tokenizer, reward_model, responses).

Trim redundant concept re-explanation (the 'Principles' list and two-phase recap duplicate the Quick start) and either fill or remove the empty 'Advanced topics' section; consider moving the full constitution templates and detailed workflows into reference files.

Dimension	Reasoning	Score
Conciseness	The body is mostly practical code, but it also re-explains concepts Claude already knows (the two-phase recap, the RLAIF/RLHF/Self-critique definitions repeated in 'Principles') and contains an empty 'Advanced topics' section, matching the 'mostly efficient but includes some unnecessary explanation' anchor 3.	3 / 5
Actionability	Concrete Python is provided throughout, but multiple examples rely on undefined helpers (create_dataset, parse_preferences, create_preference_dataset) and undefined variables (model, reward_model, responses in Workflow 3), leaving incomplete executable guidance at anchor 3 rather than the minor-gaps anchor 4.	3 / 5
Workflow Clarity	Workflows are clearly sequenced (Step 1–4), but model training is a batch/expensive operation with no validation or verification checkpoints (e.g. evaluating the reward model or harmlessness), so per the batch-operation cap workflow clarity cannot exceed anchor 3.	3 / 5
Progressive Disclosure	The skill is a single ~270-line file with decent section headers but no bundle files or one-level-deep references; substantial content (full multi-step workflows, constitution templates) that could live in separate files is inlined, fitting anchor 3.	3 / 5
	Total	12 / 20 Passed

Description

67%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 third-person, clearly states what the method does, and provides a usable 'Use for...' trigger, making it above-average. It is held back by somewhat technical trigger terms and a 'when' clause that lacks concrete user-facing trigger phrases.

Suggestions

Add concrete natural trigger phrases to the 'Use for' clause, e.g. 'Use when the user asks about safety alignment, reducing harmful or toxic outputs, RLAIF, or training harmless models without human labels'.

Broaden trigger-term coverage with synonyms a user might actually say (e.g. 'AI safety training', 'harmlessness', 'self-critique training') rather than relying on the technical 'RLAIF' jargon.

Tighten distinctiveness by naming the specific contrast (e.g. 'without human preference labels, unlike RLHF/DPO') so it is less likely to trigger for general safety-alignment requests.

Dimension	Reasoning	Score
Specificity	Names the domain and several concrete actions — 'supervised learning with self-critique/revision, then RLAIF (RL from AI Feedback)' — but stops short of comprehensive capability coverage, so it sits between anchor 3 and 5.	4 / 5
Completeness	Answers both 'what' (two-phase self-improvement training) and 'when' ('Use for safety alignment, reducing harmful outputs without human labels'), but the 'when' could be more specific with concrete trigger phrases, fitting anchor 4 rather than the fully explicit anchor 5.	4 / 5
Trigger Term Quality	Includes some relevant keywords ('safety alignment', 'harmless AI', 'RLAIF') but leans technical and lacks the common natural phrases or synonyms a user would say, matching anchor 3 rather than the fuller coverage of 4.	3 / 5
Distinctiveness Conflict Risk	Constitutional AI / RLAIF is a clear niche, yet the broad 'safety alignment' trigger risks minor overlap with adjacent RLHF/DPO safety skills, placing it at anchor 4 rather than the minimal-conflict anchor 5.	4 / 5
	Total	15 / 20 Passed

Validation

93%

Warnings & errors only

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

Validation — 15 / 16 Passed

Validation for skill structure

Criteria	Description	Result
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	15 / 16 Passed

Repository: synthetic-sciences/openscience
Path: backend/cli/skills/llm-tools/constitutional-ai/SKILL.md
Commit: 7dabbf0

Reviewed: 2 days ago

Table of Contents

Discovery Implementation Validation

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.