x-skill-improve — Session-Based Skill Alignment Analyzer

Evaluates how well an x-skill was followed during a real session, then improves the skill based on findings.

Bootstrap

0. Pin capabilities for the session per ../x-shared/capability-loading.md. The plugin.omc snapshot token (manifest key plugins.oh_my_claudecode) gates the session_search MCP tool used in step 1; degrade to JSONL-direct fallback if unavailable.

Invocation

For how to invoke skills and agents, see ../x-shared/invocation-guide.md.

Usage:

• /x-skill-improve x-do — search recent sessions for x-do usage
• /x-skill-improve x-research --since 2d — search last 2 days
• /x-skill-improve x-do /path/to/project abc123 def456 — skill + project dir + multiple sessions
• /x-skill-improve /path/to/project abc123 def456 — project dir + multiple sessions (auto-detect skill)
• /x-skill-improve abc123-def456 — single session (auto-detect skill, use current project)
• /x-skill-improve (no args) — prompt for skill name, then search

Detection

Trigger when the user:

• Asks to improve/evaluate a skill based on recent usage
• Wants to check if a skill was followed correctly
• Asks to evolve a skill based on real usage
• Says "improve skill from session", "skill alignment check", "did the skill work right"

Workflow

1. Locate Session

Use session_search (MCP tool) to find sessions where the target skill was invoked.

Parse arguments — see references/argument-parsing.md for the full parsing table and resolution rules.

Search and extract — see references/session-discovery.md for discovery queries, fallback ladder, and deep extraction parameters.

• Memory recall (only when mcp.agentmemory pinned in bootstrap-active set): one mcp__plugin_agentmemory_agentmemory__memory_sessions({ limit: 20 }) call plus one mcp__plugin_agentmemory_agentmemory__memory_smart_search({ query: "<skill name + improvement keyword>", limit: 5 }) call. When both succeed, prefer these over the ad-hoc JSONL scan for session discovery. Apply consumer rules from ../x-shared/mcp-toolbox.md § Consumer rules — drop smart-search hits where tags includes auto-import OR confidence < 0.5 before treating them as precedent. When mcp.agentmemory is not pinned, skip silently — Claude's native auto-memory file still applies and the JSONL fallback above remains the path.

2. Load Skill Files

Read the full skill directory for the identified skill. Resolve location using this precedence:

• Skill name starts with x- → check plugin source repo first. Resolve dynamically:
  1. ${X_SKILLS_PLUGIN_ROOT:-}/skills/<name>/ if env var set
  2. The directory of an x-skills git checkout (try git -C "$dir" config --get remote.origin.url | grep -q x-skills for any candidate from ~/Codes, ~/code, ~/src, $HOME)
  3. ~/.claude/plugins/cache/x-skills-marketplace/x-skills/*/skills/<name>/ (read-only — never edit here)
• Fallback for non-x skills → ~/.claude/skills/<name>/
• Plugin cache → read-only reference, never edit here

<resolved-path>/<skill-name>/
├── SKILL.md          # Always read
├── steps/            # Read all if present
├── references/       # Read all if present
├── gotchas.md        # Read if present
└── config.json       # Read if present

Build an instruction inventory — a list of every rule, gate, checklist item, and workflow step in the skill.

3. Analyze Alignment

Walk through the instruction inventory. For each item, classify based on the session:

Use the analysis rubric in references/analysis-rubric.md for skill-specific checks.

Focus on high-signal misalignments — not every trivial deviation matters. Prioritize:

• Mandatory gates that were skipped (e.g., plan review, verification)
• Steps that were worked around (signals missing guidance)
• Patterns that repeated (signals systemic issue)

4. Dual-Perspective Findings

For each misalignment, present both perspectives:

• What the skill says — quote the specific instruction
• What the session did — describe what actually happened
• Verdict:
  • UPDATE SKILL — The skill is wrong, incomplete, or too rigid. The execution was reasonable.
  • COMPLIANCE GAP — The skill is right. The execution should have followed it.
• Recommendation — Specific proposed change (file, location, content) or note for compliance

5. Present Report

Use the output format below. Then offer to apply fixes.

Output Format

Use the template in references/output-template.md.

Severity

All findings use the shared severity scale:

Applying Fixes

When the user chooses to apply:

• Resolve edit target: For x-* skills, always edit the source repo (resolved per the precedence in step 2: $X_SKILLS_PLUGIN_ROOT, then a local x-skills git checkout). Never edit the plugin cache (~/.claude/plugins/cache/) or installed copy. For standalone skills, edit ~/.claude/skills/<name>/.
• Default edit tool: native Edit / Write for all skill edits — surgical partial edits.
• UPDATE SKILL findings: Make targeted edits to the skill files. Prefer:
  • Adding exceptions to existing rules (not rewriting them)
  • Adding items to gotchas.md for newly discovered pitfalls
  • Adding missing guidance as short sections (not bloating SKILL.md)
• COMPLIANCE GAP findings: No skill change. Optionally add to gotchas.md as a reminder.
• After applying, show a summary of what changed and offer validation:

  Updates applied. Validate with /x-skill-review? [Y] Run review [N] Done

• For x-skills plugin edits: Remind user to bump version and release after fixes.

Dependencies

This skill references shared infrastructure in ../x-shared/:

• invocation-guide.md — tool invocation patterns
• severity-guide.md — finding severity scale
• workflow-chains.md — cross-skill chaining

Runtime: Requires session_search MCP tool from the oh-my-claudecode plugin. Falls back to JSONL-direct read of ~/.claude/projects/*/sessions/*.jsonl if unavailable.

Persistence

After presenting the report, append a summary line to data/alignment-log.jsonl:

{"skill":"x-bugfix","sessionId":"f7035623","date":"2026-04-01","findings":8,"updateSkill":3,"complianceGap":5,"applied":true}

This enables cross-session pattern tracking — recurring compliance gaps signal systemic issues.

• Persist lesson (only when mcp.agentmemory pinned): one mcp__plugin_agentmemory_agentmemory__memory_save({ content: "<improvement applied to skill X>", type: "lesson", concepts: "<project-slug>:x-skill-improve,<skill-name>,<improvement-class>" }) call (project-slug = basename of cwd — see ../x-shared/mcp-toolbox.md § Consumer rules). Skip silently when not pinned.

After This Skill

Updates applied? → Offer /x-skill-review on the modified skill (external user-level skill at ~/.claude/skills/x-skill-review/; if the user has not installed it, surface that fact and stop — do NOT silently no-op the slash command). Primitive: handoff (sync, depends on result). Include a handoff context block: from x-skill-improve, name of the modified skill, list of applied edits with severities, alignment-log entry written. See ../x-shared/workflow-chains.md.

Gotchas

See gotchas.md for known pitfalls — update when you encounter new ones.

Task: {{ARGUMENTS}}