CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

llm-tuning-patterns

LLM Tuning Patterns

Quality

57%

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 ./.claude/skills/llm-tuning-patterns/SKILL.md
SKILL.md
Quality
Evals
Security

Quality

Content

68%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 delivers genuinely actionable, evidence-cited parameter settings in a lean, well-sectioned format. Weaknesses are the missing validation/selection loop in the parallel-sampling workflow and a small amount of session-log filler that could be trimmed.

Suggestions

Add a validation checkpoint to the parallel sampling workflow, e.g. 'Verify each candidate proof compiles in Lean 4 before ranking; discard failures, then select the shortest passing proof.'

Remove or compress the 'Source Sessions' section — it records session history ('This session: APOLLO parity - increased max_tokens 512->4096') rather than instructing future use.

Tighten the opening: replace the two-sentence preamble with a single line or fold it into the header, letting the tables carry the guidance.

DimensionReasoningScore

Conciseness

The body is lean: parameter tables with one-line rationales and a compact copy-paste prompt. Minor trimmable padding remains — the filler sentence 'Different tasks require different LLM configurations. Use these evidence-based settings.' and the 'Source Sessions' section, which is session-log metadata rather than usable instruction. Efficient with minor over-explanation — anchor 4, not 5.

4 / 5

Actionability

Concrete, copy-paste-ready values throughout: parameter tables ('max_tokens 4096', 'temperature 0.6', 'top_p 0.95'), a complete proof-plan prompt template, and specific anti-patterns ('512 tokens truncates chain-of-thought'). The gap keeping it below 5: 'Parallel Sampling' says 'Use best-of-N selection' with no executable mechanism for how to select the best candidate.

4 / 5

Workflow Clarity

The proof workflow has a clear sequence ('Always request a proof plan before tactics'), but validation checkpoints are absent: parallel sampling (N=8-32) is a batch operation with no step for verifying which candidate actually proves the theorem or compiles. Per the batch-operation rule, workflow clarity is capped at 3 — sequence present, checkpoints missing.

3 / 5

Progressive Disclosure

The skill is self-contained with no bundle files and clear section headers (task-type tables, proof plan, sampling, anti-patterns), so all inline content is appropriately placed. The minor gap: the 'Source Sessions' section is extraneous session-log content that dilutes the otherwise clean structure — good structure with minor organization gaps, anchor 4.

4 / 5

Total

15

/

20

Passed

Description

28%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 a bare title with no actions, no trigger conditions, and no use-when guidance. It would rarely be selected by description-matching since it contains none of the phrases a user would naturally say when needing this skill.

Suggestions

Add concrete actions in third person, e.g. 'Configures LLM inference parameters (max_tokens, temperature, top_p) per task type — theorem proving, code generation, creative tasks.'

Add an explicit trigger clause: 'Use when tuning LLM parameters, choosing temperature or max_tokens settings, or configuring model sampling for proofs or code generation.'

Include natural synonyms users would say — 'model settings', 'sampling parameters', 'temperature', 'top_p' — to improve trigger matching and distinctiveness against general prompting skills.

DimensionReasoningScore

Specificity

The description is just the noun phrase 'LLM Tuning Patterns' — it names the domain but lists no actions whatsoever, not even the generic verb level of 'Processes PDF files'. It sits between anchor 1 (entirely vague) and anchor 2 (domain named, actions minimal), lifted to 2 only by the specific domain naming.

2 / 5

Completeness

There is no 'when should Claude use it' clause at all (which alone caps completeness at 3), and the 'what' is only vaguely implied by the title. This matches anchor 2: a vague 'what' and no 'when'.

2 / 5

Trigger Term Quality

The only keywords are 'LLM', 'Tuning', and 'Patterns'. Natural phrases a user would actually say — 'temperature', 'max_tokens', 'model settings', 'sampling parameters', 'top_p' — are all missing. This matches anchor 2: one or two generic keywords, missing the natural phrases users say.

2 / 5

Distinctiveness Conflict Risk

'LLM Tuning Patterns' points at a reasonably niche area (LLM parameter configuration), but as a bare title it could overlap with general prompting, model-selection, or inference-config skills. Somewhat specific but overlap risk remains — anchor 3.

3 / 5

Total

9

/

20

Passed

Validation

93%

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

CriteriaDescriptionResult

description_field

'description' is very short (19 chars), consider making it more detailed

Warning

Total

15

/

16

Passed

Repository

parcadei/Continuous-Claude-v3

Path

.claude/skills/llm-tuning-patterns/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.