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.

60

Quality

70%

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 ./backend/cli/skills/llm-tools/constitutional-ai/SKILL.md
SKILL.md
Quality
Evals
Security

Quality

Content

57%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-structured and token-efficient, with clearly sequenced two-phase workflows, but the code is peppered with undefined helper placeholders and no validation or evaluation checkpoints anywhere in the training pipelines. As a monolithic file with no reference bundle, it also inlines material that would better live in separate files.

Suggestions

Replace or define the placeholder helpers (create_dataset, parse_preferences, generate_critique, generate_revision, majority_vote) so the workflow code is executable end-to-end, and correct the TRL API usage (e.g., PPOConfig/PPOTrainer signatures).

Add validation checkpoints to both workflows: verify reward-model accuracy on held-out preference pairs before PPO, and evaluate harmlessness (e.g., red-team prompt set) after each training phase, with a fix-and-retry loop.

Remove the empty '## Advanced topics' heading and move hardware requirements and resource links into reference files (e.g., references/training.md, references/resources.md), keeping SKILL.md as a lean overview.

DimensionReasoningScore

Conciseness

The body is mostly lean code blocks and terse prompts with no re-explanation of concepts Claude already knows, matching 'efficient; minor instances of over-explanation that could be trimmed'. Not 5 because of the empty '## Advanced topics' heading, the '(no human labels needed!)' comment, and scattered filler.

4 / 5

Actionability

Real transformers/trl calls sit alongside many undefined placeholders — 'create_dataset', 'parse_preferences', 'generate_critique', 'majority_vote', 'model_1.evaluate', 'reward_model', 'tokenizer' — and questionable API shapes like 'PPOConfig(reward_model_path=...)', matching 'pseudocode instead of executable code; missing key details'. Not 4 because the snippets are not copy-paste runnable as written.

3 / 5

Workflow Clarity

Both phases have clear Step 1-4 sequences but zero validation checkpoints — no reward-model accuracy check before PPO, no post-training harmlessness evaluation — in batch training pipelines, matching 'sequence present but checkpoints missing' and hitting the batch-operation cap of 3. Not 4 because no feedback loop (validate -> fix -> retry) exists anywhere.

3 / 5

Progressive Disclosure

Section headers give structure, but there are no bundle files at all and content that belongs in separate references (full SL/RL workflow details, hardware requirements, resources) is inlined in a ~270-line body, matching 'content that should be separate is inline'. Not 4 because the empty '## Advanced topics' heading shows unfinished organization and nothing is split out.

3 / 5

Total

13

/

20

Passed

Description

83%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.

A strong description: it states both phases of the method concretely, includes an explicit 'Use for...' trigger clause, and occupies a distinct named-method niche. Minor deductions come from light marketing padding and a few missing natural synonyms.

DimensionReasoningScore

Specificity

The description names the domain and several concrete mechanisms ('supervised learning with self-critique/revision, then RLAIF', 'reducing harmful outputs without human labels'), matching the 'lists several specific actions; minor gaps' anchor. It is not 5 because 'self-improvement' and 'Powers Claude's safety system' are padded/vague and coverage stays at the phase level.

4 / 5

Completeness

It explicitly answers what ('Two-phase approach - supervised learning with self-critique/revision, then RLAIF') and when ('Use for safety alignment, reducing harmful outputs without human labels') with concrete trigger use-cases, matching the top anchor. Not 4 because the 'when' clause is explicit and specific rather than weakly implied.

5 / 5

Trigger Term Quality

Good natural keyword coverage ('safety alignment', 'harmful outputs', 'self-critique', 'RLAIF', 'harmless') that users of this domain would actually say. Not 5 because common variations like 'harmlessness', 'AI feedback', or 'constitution'-based wording are missing.

4 / 5

Distinctiveness Conflict Risk

'Anthropic's method for training harmless AI' plus RLAIF/self-critique gives a clear named-method niche with distinct triggers. Not 5 because it could still overlap with closely related alignment skills (RLHF, DPO, general safety-training skills).

4 / 5

Total

17

/

20

Passed

Validation

87%

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

Validation — 14 / 16 Passed

Validation for skill structure

CriteriaDescriptionResult

metadata_version

'metadata.version' is missing

Warning

frontmatter_unknown_keys

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

Warning

Total

14

/

16

Passed

Repository
synthetic-sciences/openscience
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.