book-sft-pipeline

Name: book-sft-pipeline
Rating: 71 (1 reviews)
Author: muratcankoylan

tessl install github:muratcankoylan/Agent-Skills-for-Context-Engineering --skill book-sft-pipeline

github.com/muratcankoylan/Agent-Skills-for-Context-Engineering

This skill should be used when the user asks to "fine-tune on books", "create SFT dataset", "train style model", "extract ePub text", or mentions style transfer, LoRA training, book segmentation, or author voice replication.

Review Score

71%

Validation Score

13/16

Implementation Score

73%

Activation Score

55%

SKILL.md

Review

Evals

Generated

12 days ago

Warnings & errors only

Criteria	Score
metadata_version	'metadata' field is not a dictionary
license_field	'license' field is missing
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata

Overall Assessment

This is a strong, actionable skill with executable code for each pipeline phase and good progressive disclosure through external references. The main weaknesses are verbosity in the 'Integration with Context Engineering Skills' section (which explains relationships Claude doesn't need) and missing intermediate validation checkpoints between pipeline phases that could catch errors earlier.

Suggestions

Remove or significantly condense the 'Integration with Context Engineering Skills' section - it explains conceptual relationships rather than providing actionable guidance
Add explicit validation checkpoints between phases, e.g., 'Verify extraction produced >50k words before proceeding to segmentation' and 'Validate chunk count is within expected range (200-400 for 80k word book)'

Dimension	Score	Reasoning
Conciseness	2/3	The skill is comprehensive but includes some unnecessary sections like 'Integration with Context Engineering Skills' which adds ~200 lines explaining how this skill relates to other skills - information Claude doesn't need to execute the pipeline. The core pipeline content is reasonably efficient.
Actionability	3/3	Provides fully executable Python code for each phase (extraction, segmentation, instruction generation, dataset building, training). Code examples are copy-paste ready with specific libraries, parameters, and expected outputs clearly defined.
Workflow Clarity	2/3	The 6-phase pipeline is clearly sequenced with a visual diagram, but validation checkpoints are only at the end (Phase 6) rather than integrated throughout. Missing explicit validation steps between phases - e.g., no 'verify extraction succeeded before segmenting' or 'validate chunk boundaries before instruction generation'.
Progressive Disclosure	3/3	Well-structured with clear sections progressing from concepts to implementation to validation. References to external files (segmentation-strategies.md, tinker-format.md, tinker.txt) are clearly signaled and one level deep. Content is appropriately split between overview and detailed references.

Overall Assessment

This description is essentially a list of trigger terms without any explanation of what the skill actually does. While it excels at providing natural keywords users might say and occupies a distinct niche, it completely fails to describe the skill's capabilities, making it impossible for Claude to understand what actions this skill enables.

Suggestions

Add a clear 'what' statement at the beginning describing concrete capabilities, e.g., 'Processes ePub files to extract and segment text, then formats content into supervised fine-tuning datasets for training style-transfer LoRA models.'
Restructure to lead with capabilities before the 'Use when...' trigger clause, following the pattern: '[What it does]. Use when [triggers].'
Specify the outputs or deliverables the skill produces (e.g., 'generates JSONL training files', 'creates conversation pairs for SFT').

Dimension	Score	Reasoning
Specificity	2/3	The description names the domain (SFT dataset creation, ePub processing, style transfer) and mentions several actions like 'fine-tune on books', 'create SFT dataset', 'extract ePub text', but doesn't comprehensively list what concrete actions the skill actually performs.
Completeness	1/3	The description only addresses 'when' (trigger conditions) but completely fails to explain 'what' the skill actually does. There's no explanation of the skill's capabilities or outputs - it's entirely composed of trigger phrases without describing functionality.
Trigger Term Quality	3/3	Excellent coverage of natural trigger terms users would say: 'fine-tune on books', 'create SFT dataset', 'train style model', 'extract ePub text', 'style transfer', 'LoRA training', 'book segmentation', 'author voice replication' - these are realistic phrases users would naturally use.
Distinctiveness Conflict Risk	3/3	The combination of ePub processing, SFT dataset creation, LoRA training, and author voice replication creates a very specific niche that is unlikely to conflict with other skills. The triggers are highly distinctive.