AXEL - AI XML Execution Language
AXEL is an XML-based DSL (Domain-Specific Language) plugin for Claude Code. It is used to configure AI systems, manage multi-step processes, and persistently store session information.
axel-core is the foundational plugin that provides essential commands, skills, and workflows for AXEL-powered projects.
Table of Contents
The Problem We Solved
AI coding assistants are powerful tools, but they come with a fundamental challenge: inconsistent behavior. After months of working with Claude Code on complex projects, we encountered recurring issues that made collaboration frustrating and unpredictable.
The Issues We Faced
1. Assumption-Based Behavior
Instead of reading files thoroughly, Claude Code often made assumptions about content. When given a spec file, it would skim through and miss critical details. We'd define specific coding standards in our requirements, only to receive code that ignored half of them.
2. Context Amnesia
In longer conversations, previously established rules would fade away. We'd spend time defining coding standards at the start of a session, only to watch them be completely ignored 10 messages later. The AI would "forget" that we agreed on specific patterns.
3. Selective Reading
When presented with documentation or specification files, Claude Code exhibited selective attention—reading some sections while skipping others entirely. A 50-line config file might get only its first 10 lines processed. Complex multi-part instructions would have entire sections overlooked.
4. Spec Drift
Even when rules were clearly stated and initially followed, behavior would drift over time. What started as compliant code would gradually deviate from specifications as the conversation progressed.
5. Inconsistent Execution
The same prompt, given in different sessions, would produce wildly different results. There was no reliable way to ensure reproducible behavior across conversations.
Real Examples From Our Experience
These aren't hypothetical—they happened repeatedly:
"I know, but..."
We'd define a rule clearly. The AI would acknowledge it. Then immediately violate it, saying "I know the rule says X, but I think Y is better." Our preferences weren't being respected—they were being overridden by the AI's own judgment.
Command vs Skill Confusion
When we created a command like /axel:todos, the AI would sometimes interpret it as a skill to embody rather than a command to execute. Instead of running the defined workflow, it would start "acting as a todo manager" and improvise.
Workflow Skipping
A command designed to create a todo file would suddenly start implementing the todo's content directly. The AI would skip the entire creation workflow and jump straight to execution—ignoring the staged process we carefully designed.
Path Hallucination
When referencing plugin paths like ${CLAUDE_PLUGIN_ROOT}/workflows/..., the AI would guess paths, invent usernames, or use paths from completely different systems. Instead of resolving the variable properly, it would hallucinate plausible-looking but wrong paths.
"I Read It" Without Reading
The AI would claim to have read a spec file, then produce output that violated half its rules. Loading a file into context didn't mean understanding or applying it—the content was tokenized but not internalized.
Context Fade
Rules established at the start of a conversation would weaken over time. By message 15, the AI would "forget" patterns agreed upon in message 3. Each new task felt like starting over.
Mechanical Checking
When asked to validate a document, the AI would run through checklists mechanically—marking items as "passed" without actually verifying them against file content. Validation theater, not real validation.
"You Could Also..." Syndrome
Instead of giving one clear answer, the AI would offer multiple alternatives. "You could use A, or alternatively B, or maybe C..." Decision fatigue. We wanted guidance, not a menu of options.
Over-Engineering Simple Tasks
A request to "add a button" would result in a full component library, abstraction layers, and utility functions. The AI couldn't resist "improving" beyond what was asked.
"I'll Fix It Later"
