A Claude skill that brings structured, senior-engineer thinking to IT problem investigation. Built on the OHTV-D framework — Observe, Hypothesize, Test, Verify, Document — the same pattern used by SREs and infrastructure engineers in production environments.
Includes integrated token compression based on the caveman philosophy, reducing output length by approximately 65% while preserving full technical accuracy.
Most AI-assisted debugging jumps straight to solutions before understanding the problem. This skill enforces a disciplined investigation process: gather facts first, form ranked hypotheses second, test with targeted commands third. It prevents the pattern of trying random fixes and hoping one works.
The skill is domain-agnostic and has been tested across:
- Application development (Python, Node.js, Java, and others)
- Infrastructure and DevOps (Linux, Docker, Kubernetes)
- Databases (MySQL, PostgreSQL, Redis)
- Networking (DNS, TLS, firewall, connectivity)
- Security incident triage
- General IT support
Every session follows five phases:
Observe — Separate facts from assumptions. Identify scope, timeline, recent changes, and available artifacts (logs, traces, error codes).
Hypothesize — Form two to four ranked hypotheses based on evidence. Draws from a taxonomy of universal failure patterns: recent changes, resource exhaustion, dependency failure, race conditions, data corruption, environment drift, and capacity limits.
Test — Provide specific, executable steps to verify each hypothesis. Ordered from least invasive to most. Evidence is preserved before any system changes.
Verify — Confirm the root cause is identified, the fix resolves the original symptom, and no side effects were introduced.
Document — Produce a structured incident report with timeline, root cause, solution applied, and prevention actions.
Activated with /caveman. Compresses Claude's responses while keeping all technical content exact — code, commands, error strings, and API names are never abbreviated.
/caveman full mode (default)
/caveman lite drop filler, keep full grammar
/caveman ultra maximum compression, fragment style
stop caveman return to normal prose
The compression automatically disables for security warnings, irreversible actions, and any step where ambiguity could cause harm.
Example of the same diagnosis at different verbosity levels:
Normal: "The reason your service is not responding is most likely because
the connection pool has been exhausted. I would recommend checking
the current connection count and increasing the pool size."
Full: "Connection pool exhausted. Check active conns, increase pool_size
in config."
Ultra: "Pool full. Check conns → up pool_size."
it-problem-solver/
├── SKILL.md
└── references/
├── diagnostic-commands.md
└── error-pattern-library.md
SKILL.md contains the main framework. The two reference files are loaded on demand — diagnostic commands when specific platform commands are needed, error patterns when matching a known error signature.
Claude.ai
Download it-problem-solver.skill from the releases page. Open Claude.ai, go to Settings, then Skills, and upload the file.
Claude Code
git clone https://github.com/YOUR_USERNAME/it-problem-solver
cp -r it-problem-solver ~/.claude/skills/diagnostic-commands.md covers Linux, Docker, Kubernetes, MySQL, PostgreSQL, Redis, Nginx, Apache, and Windows — roughly 50 commands organized by platform.
error-pattern-library.md covers Python, JavaScript/Node.js, Java, MySQL, PostgreSQL, Docker, Kubernetes, networking, Linux, and Git — roughly 60 named error patterns with likely causes and first investigation steps.
Token compression approach adapted from caveman by Julius Brussee (MIT).
OHTV-D framework adapted from SRE and incident response best practices.
MIT