A Control Surface Map
Each surface — instruction files, naming, tool/skill descriptions, hooks, context, MCP — whether it's present, its strength, and whether it currently acts as suggestion or instruction.
Diagnostic · the scan
See where your setup is making a suggestion, not an instruction.
A scan of your Claude Code setup, read against the Control Surfaces taxonomy — the named failure modes that make an agent quietly ignore the rules you thought you'd set.
It tells you which are present, how severe, and exactly where — by name, scored, against your own code.
Request a scan →
By enquiry at launch (POA).
What it is
The Masterclass framework, applied to a real repo.
The Diagnostic checks a Claude Code setup for the taxonomy's named patterns — the recurring ways an instruction reads as binding while nothing actually stops a violation.
You get a report, not a checklist score: what's already working, and every gap named, severity-scored, and located in your own code.
What it looks for
The named failure modes.
It checks for the taxonomy's named patterns. Among them:
-
Critical Rules in Prose
A zero-tolerance rule (“MUST”, “NEVER”) written where prose can only suggest — so it reads as binding while nothing stops a violation.
-
Non-Enforcing Hook
A hook exists, so the setup looks enforced — but it doesn't gate (exits zero, auto-fixes instead of blocking, or fails open).
-
Generic Naming / Low Semantic Match
An instruction, skill, or tool whose name doesn't match the task — so the agent passes over it.
-
Competing Instruction Surfaces
Two or more instruction files with no hierarchy, so the agent may read one, none, or the wrong one.
-
Reference Fragmentation
Files pointed to in prose instead of pulled into context — so whether they arrive is left to the agent's discretion.
-
No Control Surface
Claude Code run on a real codebase with no instruction or enforcement surface at all.
-
MCP Surface Overload
Many bridged tools competing for the agent's selection attention.
-
Context / Attention Dilution
Non-actionable text crowding the real instruction out of the window.
…and the rest of the taxonomy — Session Invariants, Manual Context Loading, Compliance Illusion, No Universal Config.
What you get
A report that names, scores, and locates every gap.
Strengths
What's already working in your setup, cited to specific files.
Severity-scored gaps
Each named anti-pattern present, scored HIGH / MEDIUM / LOW, with the exact file and the offending lines as evidence.
What it deliberately withholds — twice over
You get the findings, not the recipe or the machine.
Free OSS tools cover Anthropic best-practice checks only; the full proprietary taxonomy is the paid scan.
The fix
The Diagnostic names and scores what's wrong; the remediation method — how to fix it and verify the fix holds — transfers in the Masterclass and the Workshop.
The method
How the scan detects and scores is not disclosed — measurement method patent-pending. You receive the findings, not the machine that produced them.
Who's behind it
Philip Forshaw
Olinave · ex-Apple AI Strategy & Operations
Ex-Apple AI Strategy & Operations lead, with decades of training and delivery behind him. The framework is patent-pending.
The method is a named taxonomy of the recurring failure modes, grounded in controlled experiments and Anthropic's own published guidance — then measured across 300+ real Claude Code setups. Not theory; the scan reads your setup against the evidence.
Get a full diagnostic scan
Point it at your repo.
By enquiry— public-repo or a private scan via secure upload. Tell us which and we'll scope it.
Request a quote →Public-repo or private · POA