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

68%

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 ./skills/safety-alignment/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 body is well-organized into quick-start and per-phase workflows with mostly executable-looking code, but the code relies on undefined helpers/globals, omits validation checkpoints for destructive training runs, and points to reference files that are not present in the bundle.

Suggestions

Resolve or remove the dangling references: create references/constitution-design.md, rlaif-comparison.md, and cot-critique.md, or drop the links so navigation is not broken.

Make the code copy-paste runnable by defining or importing the helpers it depends on (create_dataset, parse_preferences, majority_vote, and the model/tokenizer/reward_model/CONSTITUTION globals), or explicitly mark them as placeholders to fill in.

Add validation/evaluation checkpoints to the training workflows (e.g. evaluate the reward model on held-out preferences before PPO, sanity-check revised responses before SFT) so the destructive/batch steps have feedback loops.

Dimension	Reasoning	Score
Conciseness	Mostly efficient with code-forward workflows, but padded with restated explanations (the constitution appears three times, 'Key concept' prose, a 'no human labels needed!' comment) that could be trimmed.	3 / 5
Actionability	Code looks concrete but is not executable: undefined helpers (create_dataset, parse_preferences), undefined globals (model, tokenizer, reward_model, CONSTITUTION), and an invented RewardConfig.reward_model_path field, so steps are pseudo-runnable rather than copy-paste ready.	3 / 5
Workflow Clarity	Steps are clearly sequenced per phase, but the batch/destructive training operations (SFT, reward training, PPO) have no validation or verification checkpoints and no eval feedback loop, so the cap applies.	3 / 5
Progressive Disclosure	Structure is reasonable with three clearly signaled reference files, but the references/ directory does not exist in the bundle, so those links are dead and navigation is broken rather than one-level-deep and resolvable.	3 / 5
	Total	12 / 20 Passed

Description

87%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 specific, complete, and distinct, naming the two phases and giving an explicit 'Use for' trigger clause. It could add a few more natural synonyms (e.g. 'AI safety training', 'red-teaming') to round out trigger coverage.

Dimension	Reasoning	Score
Specificity	Names the domain and several concrete actions ('supervised learning with self-critique/revision', 'RLAIF (RL from AI Feedback)', 'reducing harmful outputs'), with only minor gaps in coverage versus the comprehensive 5-anchor.	4 / 5
Completeness	Explicitly answers both what ('Anthropic's method for training harmless AI through self-improvement... Two-phase approach') and when ('Use for safety alignment, reducing harmful outputs without human labels') with a concrete trigger clause.	5 / 5
Trigger Term Quality	Good natural keywords ('safety alignment', 'harmless', 'RLAIF', 'self-critique', 'human labels') that users would plausibly say, though common synonyms like 'AI safety training' or 'red-teaming' are missing.	4 / 5
Distinctiveness Conflict Risk	A clear niche (Constitutional AI / RLAIF) with distinct triggers that would not fire for adjacent RLHF/DPO or runtime-moderation skills.	5 / 5
	Total	18 / 20 Passed

Validation

81%

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 — 13 / 16 Passed

Validation for skill structure

Criteria	Description	Result
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning
relative_links	Relative link issues: 3 missing	Warning
referenced_paths_exist	Referenced path issues: 6 missing	Warning

	Total	13 / 16 Passed

Repository: OpenLAIR/dr-claw
Path: skills/safety-alignment/constitutional-ai/SKILL.md
Commit: 4b35756

Reviewed: 1 day 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.