CtrlK
BlogDocsLog inGet started
Tessl Logo

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.

52

Quality

58%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

SecuritybySnyk

Passed

No findings from the security scan

Fix and improve this skill with Tessl

tessl review fix ./07-safety-alignment/constitutional-ai/SKILL.md
SKILL.md
Quality
Evals
Security

Quality

Content

50%

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

The skill body is well-organized with clear workflows and concrete library code, but is held back by undefined helper functions that break executability, absent validation checkpoints in the training workflows, and three referenced files that do not exist on disk.

Suggestions

Define or stub the helper functions (create_dataset, parse_preferences, create_preference_dataset, majority_vote) so the code is executable end-to-end, or note them as user-supplied.

Add validation/feedback checkpoints in the workflows — e.g., evaluate the reward model's accuracy before PPO, and verify revised-response quality before fine-tuning.

Create the missing references/ files (constitution-design.md, rlaif-comparison.md, cot-critique.md) or remove the dead links so progressive disclosure navigation actually resolves.

DimensionReasoningScore

Conciseness

The body is mostly efficient and action-oriented with code, but includes mild over-explanation (a 'Key concept' gloss, a re-explained constitution, and restated principles) that could be tightened. Not a 3 because some tokens restate what Claude already knows; not a 1 because it is not padded with basic concept explanations.

2 / 3

Actionability

Provides substantial concrete Python (SFTTrainer, RewardTrainer, PPOTrainer, pipeline), but several steps rely on undefined helper functions — create_dataset(), parse_preferences(), create_preference_dataset(), majority_vote(), model.evaluate() — making the code non-copy-paste-executable. Not a 3 because of these incomplete pseudocode gaps; not a 1 because real, specific library calls are present.

2 / 3

Workflow Clarity

Workflows are clearly sequenced into numbered steps (SL phase, RL phase), but there are no validation checkpoints or feedback loops for costly batch/training operations (e.g., verifying reward-model quality before PPO, or checking critique quality before fine-tuning). Per the rubric, missing validation in these contexts caps workflow clarity at 2; not a 1 because the sequence is explicit.

2 / 3

Progressive Disclosure

The body signals three one-level-deep references (constitution-design.md, rlaif-comparison.md, cot-critique.md), but the references/ directory does not exist, so the links are dead. Per the rubric, progressive disclosure should be scored against the actual bundle structure, which here has structure but broken navigation; not a 1 because the overview is organized and references are shallow, not a deeply nested chain.

2 / 3

Total

8

/

12

Passed

Description

67%

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 and well-differentiated, naming concrete training actions and a distinct niche. Its main weakness is the when-to-use guidance, which states purpose rather than offering natural-language triggers, capping completeness and trigger quality.

Suggestions

Replace 'Use for safety alignment, reducing harmful outputs without human labels' with an explicit 'Use when...' trigger clause using natural user phrasing (e.g., 'Use when aligning a model for safety, reducing harmful or toxic outputs without human labels, or training with AI-generated feedback').

Add common user-facing keyword variations alongside the jargon (e.g., 'safety training', 'harmlessness', 'AI feedback preferences') so trigger terms read like what a user would actually say.

DimensionReasoningScore

Specificity

Lists multiple concrete actions — 'training harmless AI through self-improvement', 'supervised learning with self-critique/revision', then 'RLAIF (RL from AI Feedback)', and 'reducing harmful outputs without human labels' — matching the 'lists multiple specific concrete actions' anchor. Not a 2 because the actions are specific and enumerated rather than a single domain mention.

3 / 3

Completeness

Clearly answers 'what does this do' but the only when-guidance is 'Use for safety alignment, reducing harmful outputs without human labels', which states purpose rather than a natural 'Use when...' trigger clause. Per the rubric, a missing explicit trigger clause caps completeness at 2; not a 1 because the what is strong and a use-purpose is present.

2 / 3

Trigger Term Quality

Contains relevant terms like 'safety alignment', 'self-critique', and 'harmless', but leans heavily on technical jargon such as 'RLAIF (RL from AI Feedback)' with missing common user-facing variations. Not a 3 because the natural-term coverage is incomplete; not a 1 because some natural keywords ('safety alignment') are present.

2 / 3

Distinctiveness Conflict Risk

The description carves a clear niche (Constitutional AI / RLAIF for harmlessness) with distinct triggers unlikely to overlap with unrelated skills. Not a 2 because the domain is specific and clearly bounded.

3 / 3

Total

10

/

12

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.

Validation13 / 16 Passed

Validation for skill structure

CriteriaDescriptionResult

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
Orchestra-Research/AI-Research-SKILLs
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.