Glossary
What the words on this site mean
tokencheat uses a few terms of its own. Each one is defined here in plain language, with a link to where you use it.
- Instruction file
A text file your AI coding assistant reads before every request, such as CLAUDE.md, AGENTS.md or .cursorrules.
Because it is sent with every request, every line costs tokens on every turn, and stale or contradictory lines change what the agent does.
Where you use it →- Always-on context
Everything an agent loads on every turn whether the task needs it or not: instruction files, skill descriptions and MCP tool schemas.
Where you use it →- Cheat Code
tokencheat's name for one reviewed rulebook — how you work, which technologies are deliberate, and what an agent may do without asking — exported to every agent you use.
You build it once in the browser and export host files for Claude Code, Cursor, Copilot and AGENTS.md-compatible tools, so every agent follows the same policy.
Where you use it →- Stage Clear
The Cheat Code builder's word for finished: the draft has been reviewed and approved, which unlocks its exports.
Where you use it →- Mission Brief
The human-readable policy document a Cheat Code compiles to — the version a person reads rather than an agent.
Where you use it →- Context Pack
A bundle of rules that loads only for a certain kind of task, instead of riding along on every request.
Where you use it →- Save State
A saved version of your Cheat Code. You can compare any two and revert, and a revert adds a new version rather than deleting history.
Where you use it →- Policy drift
When the instructions, packages or tools in a project stop matching the policy you decided on.
Where you use it →- IAS (Instruction Architecture Score)
A 0–100 diagnostic of how an instruction set is organised, not a promise of savings.
It combines always-on efficiency, duplication, progressive disclosure, trigger specificity, cohesion, portability, freshness, tool overhead, provenance and validation coverage.
Where you use it →- Runtime / Policy / Capability
The three layers tokencheat splits instructions into: session mechanics, cross-cutting rules, and the domain expertise that makes a skill valuable.
Runtime and policy are optimised aggressively; capability is protected, because shortening it removes what the agent knows how to do.
Where you use it →- Progressive disclosure
Keeping a short summary always loaded and letting the agent open the full instructions only when a task needs them.
Where you use it →- Optimization levels (L0–L4)
How much a build is allowed to change: L0 audits only, L1 removes exact duplicates, L2 moves material out of always-on context, L3 compresses wording and needs evaluation, L4 deliberately drops capability.
Where you use it →- Skill (SKILL.md)
A folder of instructions an agent loads on demand for one kind of job, described by a short summary that is always visible.
Where you use it →- MCP server
A plug-in that gives an agent extra tools. Each enabled server adds its tool descriptions to the always-on context.
Where you use it →- Token
The unit AI models read and bill by — roughly three to four characters of English text, and fewer for code or non-English text.
Where you use it →