tokencheat Skill Compiler
Audit and optimize AI agent skills without blindly removing capability.
The Skill Compiler treats instructions like a build system: measure what is always active, classify Runtime / Policy / Capability, identify duplicated or deferable context, package portable outputs, and require evidence before claiming an optimized build preserves behavior.
Live now — Prompt → Skill
Move reusable workflows out of always-on context.
Package a reusable prompt as a portable Agent Skill. The browser builder validates the artifact, measures catalog versus activated context, and exports the generated skill locally.
Turn a reusable prompt into a portable skill
tokencheat packages your instructions into a deterministic Agent Skill artifact with validation, progressive-disclosure metrics, and a portable archive. This builder runs locally in the browser.
Optional portability metadata
What the full compiler does
Audit the architecture before changing the words.
The live Prompt → Skill builder is the first public compiler workflow. The broader Skill Compiler is designed to inspect whole instruction systems, identify what must remain active, what can move on demand, and what can be structurally shared without blindly deleting capability.
Report anatomy
A report should show what changed, why, and how certain we are.
Estimated context, provider-measured usage, and behavioral evaluation are separate evidence classes. tokencheat does not convert one into another in marketing copy.
Instruction graph
Model runtime, policy, capability, examples, references, tools, outputs, safety constraints and dependencies as structured nodes instead of editing Markdown blindly.
Progressive disclosure
Classify material as always-active, on-demand, reference, rare-path or do-not-move based on need frequency, failure cost and retrievability.
Semantic criticality
Lock destructive-operation rules, security boundaries, mandatory tests, approval gates and downstream output contracts unless evaluation supports a change.
Portable by design
One internal representation. Multiple host packages.
tokencheat should not make its internal model depend on any one vendor format. Host and plugin standards are export adapters around a neutral instruction representation.
Provenance first
Pin source repository, commit, path, compiler version, build profile and transformation ancestry.
License-aware output
Unknown or restrictive licenses can still be analyzed locally, but public derivative distribution stays blocked until eligibility is established.
Parity before “Optimized”
A compiled artifact earns a TC Optimized label only after a defined behavioral-parity threshold; otherwise it remains Experimental.
Upstream vs. tokencheat build
Use the same model, repo state, user task, host, tool permissions and environment. Measure input/output/cache tokens, tool calls, task success, constraints, retries, latency, rubric score and failure modes.
Where skills get their context
A skill still needs a policy around it.
The Skill Compiler packages a reusable workflow. A Cheat Code is the standing policy that workflow runs inside — evidence standards, technology choices, and which actions need confirmation before an agent takes them.
Founding waitlist
Keep this analysis. The workspace saves it.
The paid workspace adds saved reports and score history, API and CLI access, and team governance — opening to the founding waitlist first, at founding pricing locked in before public launch. This tool stays free either way.