Update Training Materials Skill

Audit training materials against the official Claude Code changelog to find stale references, missing feature coverage, and legacy content. Optionally fix all issues in parallel.

Usage

code

/update-training                          # Scan only (audit report, no changes)
/update-training --fix                    # Scan and fix all issues
/update-training --fix --severity error   # Scan and fix only errors
/update-training --dry-run                # Show what would be fixed without applying
/update-training 2.0.0-2.1.37            # Audit only changes in version range
/update-training --focus docs             # Limit scan to docs/ only
/update-training --focus examples        # Limit scan to examples/ only

Arguments

Argument	Description
`[version-range]`	Filter changelog to specific version range (e.g., `2.0.0-2.1.37`)
`--fix`	Apply fixes after scanning (without this, only reports findings)
`--severity <level>`	Filter fixes by minimum severity: `error`, `warning`, or `info` (default: all)
`--dry-run`	Show what would be fixed without writing any files
`--focus <area>`	Limit scan scope: `docs`, `examples`, or `all` (default: all)

Default behavior: Scan only — display audit report without modifying files. --fix is required to apply changes.

Architecture

Three specialized agents work in phases:

Agent	Purpose	Count
changelog-auditor	Fetch and parse official changelog into structured JSON	1
training-scanner	Scan assigned file group against changelog data	10 (parallel)
training-fixer	Apply targeted fixes based on scanner findings	10 (parallel)

Orchestration Flow

code

/update-training [version-range] [--fix] [--severity <level>] [--dry-run] [--focus <area>]
       │
       ▼
  1. PARSE: Extract arguments and flags
       │
       ▼
  2. FETCH: Spawn 1x changelog-auditor agent
       │    - Fetches official CHANGELOG.md
       │    - Parses into structured JSON
       │    - Filters by version-range if specified
       │    - Output: {versions[], stale_patterns[], new_features[]}
       │
       ▼
  3. DISCOVER: Glob training files based on --focus flag
       │
       ▼
  4. GROUP: Split files into 10 groups by directory/topic proximity
       │
       ▼
  5. SCAN: Spawn 10x training-scanner agents IN PARALLEL
       │    - Input: assigned file group + parsed changelog
       │    - Output: JSON findings array per agent
       │
       ▼
  6. CONSOLIDATE: Merge and deduplicate all scanner findings
       │
       ├── No --fix → Exit with audit report
       │
       └── --fix →
       │
       ▼
  7. FILTER: Apply --severity filter to findings
       │
       ▼
  8. FIX: Spawn 10x training-fixer agents IN PARALLEL
       │    - If --dry-run: report what would change
       │    - Otherwise: apply targeted edits
       │
       ▼
  9. DISPLAY: Show fix summary with counts

Instructions

Step 1: Parse Arguments

Extract from the user's command:

•version_range: positional argument matching pattern N.N.N-N.N.N (or null for all versions)
•fix_mode: true if --fix is present
•severity_filter: value after --severity (or null to include all severities)
•dry_run: true if --dry-run is present
•focus_area: value after --focus (or all)

If --dry-run is present, implicitly enable fix_mode (dry run is a variant of fix).

Step 2: Fetch and Parse Changelog

Spawn 1 changelog-auditor agent.

Use subagent_type: "changelog-auditor" with the Task tool.

Agent Prompt:

code

Audit the official Claude Code changelog.

Version range filter: {version_range or "all versions"}

Fetch the changelog from:
https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md

Parse ALL entries and return structured JSON with:
- versions[] - all version entries with their changes
- stale_patterns[] - outdated references with replacements
- new_features[] - significant new features for training coverage
- deprecations[] - deprecated features
- removals[] - removed features
- renames[] - renamed features/paths
- breaking_changes[] - breaking changes

Output the full JSON object as specified in your agent instructions.

Wait for the agent to complete. Extract the parsed changelog JSON from its response.

Step 3: Discover Training Files

Use Glob to find training files based on the focus_area:

Focus Area	Glob Patterns
`all` (default)	`docs/.md`, `examples//.md`
`docs`	`docs/*.md`
`examples`	`examples/*/.md`

Collect all matching file paths.

Step 4: Split Files into 10 Groups

Split discovered files into exactly 10 groups using directory/topic proximity:

Grouping Algorithm:

•Sort files by directory then filename to cluster related files
•Group by directory first: files in the same directory stay together
•Distribute evenly: aim for roughly equal file counts per group
•Name each group based on the primary directory/topic of its files

If fewer than 10 files exist, use fewer groups (one file per group minimum).

Group Structure:

json

{
  "groups": [
    {
      "id": 1,
      "name": "core-docs-01-10",
      "files": ["docs/01-what-is-claude-code.md", "docs/02-installation.md", ...]
    },
    {
      "id": 2,
      "name": "core-docs-11-20",
      "files": ["docs/11-hooks.md", "docs/12-permissions.md", ...]
    }
  ]
}

Step 5: Spawn Parallel Scanner Agents

CRITICAL: Spawn all 10 scanner agents in a SINGLE message with 10 Task tool calls.

Use subagent_type: "training-scanner" for each.

Agent Prompt Template:

code

Scan training materials in GROUP {N}: {group_name}

ASSIGNED FILES:
{list of file paths, one per line}

CHANGELOG DATA:
{full parsed changelog JSON from Step 2}

Scan each file for:
1. STALE references - outdated patterns from stale_patterns list
2. MISSING coverage - new features from new_features list not covered where relevant
3. LEGACY content - removed/deprecated features still presented as current

Output JSON with findings array as specified in your agent instructions.

Wait for ALL scanner agents to complete. Collect all responses.

Step 6: Consolidate Scanner Results

After all scanner agents complete, merge and deduplicate findings:

•Collect: Extract the JSON findings from each agent response
•Concatenate: Merge all findings arrays into one
•Deduplicate: Remove duplicates by (file, line, description) tuple, keeping highest severity
•Re-number: Assign sequential IDs per type (STALE-001, STALE-002, etc.)
•Sort: By severity (error > warning > info), then file path, then line number

Track consolidation stats:

•raw_findings: Total findings before dedup
•after_dedup: Final deduplicated count
•dedup_rate: Percentage reduction

Step 7: Display Audit Report

Display the consolidated results:

code

Training Materials Audit Report
================================

Scan Summary:
  Files scanned: 45
  Groups: 10
  Scanner agents: 10

Consolidation:
  Raw findings: 80
  After dedup: 35 (56% dedup rate)

Summary:
  Total findings: 35
  Errors: 5
  Warnings: 20
  Info: 10

By Type:
  Stale references: 12
  Missing coverage: 15
  Legacy content: 8

Findings:
---------
[ERROR] STALE-001 (docs/02-installation.md:35)
  References ~/.claude.json which was moved to ~/.claude/settings.json
  Evidence: "Edit your `~/.claude.json` to configure..."
  Fix: Replace `~/.claude.json` with `~/.claude/settings.json`

[ERROR] LEGACY-001 (docs/10-settings.md:42)
  Shows 'claude config' command which was deprecated in v1.0.7
  Evidence: "Run `claude config set theme dark` to change..."
  Fix: Replace with direct settings.json editing instructions

[WARNING] MISSING-001 (docs/05-hooks.md)
  New hook event 'notification' not covered in hooks documentation
  Fix: Add section covering the notification hook event added in v2.1.0
...

If --fix is NOT set: Stop here. Display the report and exit.

If --fix IS set: Continue to Step 8.

Step 8: Filter Findings by Severity

Apply the --severity filter:

`--severity` Value	Include
`error`	Only severity = "error"
`warning`	severity = "error" OR "warning"
`info` or not set	ALL findings

Group filtered findings by file, then reassign to the original 10 groups for parallel fixing.

Step 9: Spawn Parallel Fixer Agents

CRITICAL: Spawn all fixer agents in a SINGLE message with multiple Task tool calls.

Only spawn fixer agents for groups that have findings after filtering. Use subagent_type: "training-fixer" for each.

Agent Prompt Template:

code

Fix training material issues in GROUP {N}: {group_name}

DRY RUN: {true|false}

FINDINGS TO FIX:
{JSON array of findings for this group's files}

For each finding:
1. Read the file and verify the issue still exists
2. Apply the suggested fix with minimal changes
3. Record the result (applied or skipped with reason)

{If dry run: "DO NOT apply any changes. Only report what would be changed."}

Output JSON with fixes_applied and fixes_skipped arrays as specified in your agent instructions.

Wait for ALL fixer agents to complete. Collect all responses.

Step 10: Display Fix Summary

code

Fix Summary
===========

{If dry run: "DRY RUN - No changes were applied"}

Applied ({count}):
  STALE-001: Replaced ~/.claude.json with ~/.claude/settings.json in docs/02-installation.md
  LEGACY-001: Updated 'claude config' to settings.json editing in docs/10-settings.md
  ...

Skipped ({count}):
  MISSING-001: No clear section to add notification hook coverage (docs/05-hooks.md)
  ...

Totals:
  Findings processed: 35
  Fixes applied: 28
  Fixes skipped: 7

Error Handling

•If the changelog fetch fails, report the error and stop
•If any scanner agent fails, continue with results from others and note the failure
•If any fixer agent fails, continue with others and report failures
•If an agent returns invalid JSON, skip that agent's output and continue
•If grouping produces fewer than 10 groups (not enough files), use as many groups as there are files
•Always display partial results even if some agents fail