README.md

Customization Guide

This plugin is designed to work out-of-the-box with sensible defaults. This document covers tuning for specific environments and use cases.

---

Hook Customization

Hooks live in hooks/hooks.json. Both hooks use prompt-based logic — customization means editing the prompt text.

Adjusting Prompt Architect Clarity Threshold

The default threshold is 7/10. Prompts scoring ≥7 pass through silently.

Lower the threshold (more permissive — fewer interruptions): Find in hooks/hooks.json (UserPromptSubmit prompt):

Score ≥7 → approve.

Change to:

Score ≥6 → approve.

Use this if you find the hook interrupting too many prompts you consider clear enough.

Raise the threshold (stricter — more clarifications):

Score ≥8 → approve.

Use this if you frequently find yourself needing to re-run tasks after missed context.

Disabling a Hook Entirely

To disable the Prompt Architect but keep the MLOps Guard:

{
  "UserPromptSubmit": [],
  "PreToolUse": [ ... ]
}

To disable both hooks, remove or empty both arrays.

CI/CD and Automated Sessions

In automated pipelines where no human is present to answer clarifying questions, hooks should not fire. Add this to the top of each hook prompt to enable a bypass:

If the environment variable CLAUDE_SKIP_HOOKS is set to "1", output {"decision":"approve"} immediately.

Then in your CI environment:

export CLAUDE_SKIP_HOOKS=1

Extending Domain Keywords

To add more keywords for domain classification, edit the DOMAIN CLASSIFY section of the UserPromptSubmit prompt. Example — adding reinforcement learning and RL to ML_RESEARCH:

• ML_RESEARCH — model, train, fine-tune, dataset, eval, paper, embedding, loss, architecture, rl, reinforcement

Keep keywords short (1-2 words) and lowercase. Longer phrases slow down classification.

MLOps Guard — Warn-Only Mode

By default, the MLOps Guard uses ask_user which pauses execution. To make it log-only (never interrupt), change:

"decision": "ask_user"

to:

"decision": "approve"

and remove the quality gate response, keeping only the approve path. This turns the guard into a silent observer — useful for teams that want visibility without interruption.

Adding Custom MLOps Checks

To add a 5th check (e.g., type hints required), add to the AUDIT section:

5. TYPE_HINTS: function definitions include ': ' type annotations or 'def.*->.*:' return type

And update the pass threshold logic accordingly.

---

Skill Customization

Skills are Markdown files. Edit them directly to tune for your stack.

Changing Default Framework References

In skills/mlops-standards/SKILL.md and skills/mlops-standards/references/tracking-patterns.md, references default to PyTorch + W&B + MLflow. To switch defaults to JAX + Neptune:

1. Find torch.manual_seed references → add jax.random.PRNGKey equivalents
2. Find wandb.init references → add neptune.init equivalents
3. Update the W&B vs. MLflow decision table to include Neptune

Adding Domain-Specific Patterns to Prompt Architect

In skills/prompt-architect/references/domain-patterns.md, add new patterns following the existing format:

### ML7: My Custom Pattern
Signal: [how to recognize it]
Fix: [specific fix]
Example: [before → after]

---

Command Customization

Commands are Markdown files with YAML frontmatter. Common adjustments:

Change Default Tracking Backend

In commands/experiment.md, find:

- Tracking backend: W&B (default) / MLflow / both

Change the default to match your infrastructure.

Restrict /mlops Output to Your Cloud Provider

Add to the beginning of commands/mlops.md:

Default cloud provider: AWS. Use AWS-native services (ECS, ECR, CloudWatch, SageMaker)
unless the user specifies otherwise.

Add Mandatory Headers to Generated Code

To enforce copyright headers on all generated Python files, add to commands/experiment.md Quality Standards section:

Every generated .py file must begin with:
# Copyright (c) [year] TechKnowmad AI. All rights reserved.

---

Version Pinning

When you customize this plugin, note the base version in your fork's CHANGELOG so you can track upstream changes that may affect your customizations.