Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/xxyoudeadpunkxx/sensecraftxstudio/llms.txt

Use this file to discover all available pages before exploring further.

An uncontracted AI session has no fixed frame: the assistant constructs its plan from the first plausible interpretation, expands local work silently, and reports activity as completion. A contracted session — one where AGENTS.md is the active operative frame — introduces deliberate friction at each point where unjustified confidence would otherwise enter. The assistant must make the real object, authority, scope, and epistemic state visible before and after every consequential move.

​
The Session Lifecycle

1

Activate

Before giving any task, give the assistant the activation instruction:
Read AGENTS.md in full and use it as the operative frame before acting on this workspace.
Activation means AGENTS.md governs how the next response is read, framed, and approached — not only what actions are permitted. This instruction must come first, in the same session turn or immediately before the task.
2

Frame the task

Before acting, the assistant closes context along three axes:
  • Object — the real target being inspected or changed, distinct from the stated request
  • Authority — who or what authorizes the move; not inferred from filename, freshness, or tone
  • Mode — the single operating mode the task requires
The assistant must also identify the destination of any output, the evidence needed to proceed, and any plausible alternative interpretations of the task.
3

Read the move

Before executing, the assistant determines whether the proposed move stays contained or expands scope, risk, or structure.
  • Contained move → proceed
  • Move that expands scope, risk, or structure → surface it before proceeding
Apparent locality does not prove contained consequence. The assistant reads enough of the surrounding system to judge whether the move remains inside the requested scope.
4

Execute minimally

Once the move is justified, the assistant uses the smallest correct intervention:
  • the smallest correct read
  • the smallest correct change
  • no opportunistic restructuring, cleanup, or abstraction
It does not silently transform a bug fix into a cleanup, a cleanup into architecture work, or a local preference into repository policy.
5

Verify and close

Before marking the task converged, the assistant confirms that the live surface — rendered output, active preview, open file, or deployed artifact — reflects the current source state. A stale cache, mismatched branch, or artifact built from an older version is a hold condition, not a minor detail.
6

Final response

For any response that includes a consequential change, recommendation, or finding, the assistant closes with four compact fields:
  • Touch — what changed and what remained untouched
  • Ground — the files, evidence, rules, or verified observations supporting the move
  • State — what is verified, inferred, unresolved, or not inspected
  • Convergence — whether the task is converged, open, blocked, or divergent
These fields must reflect the actual task state. They should not be reconstructed from memory only after the work is finished.

​
Contract Drift

The contract has drifted when AGENTS.md is no longer the active operative frame governing the current response.
Drift is present when the contract has become any of the following:
  • A one-time reading — the assistant read AGENTS.md at the start of the session and has since stopped applying it
  • Memory-only guidance — the contract exists in memory but does not shape how the next move is chosen or reported
  • Background context with no effect — AGENTS.md is acknowledged but not reflected in execution
  • Subordinate to nearby file assumptions — the assistant treats inferences drawn from adjacent files as overriding the operative contract
  • A rhetorical reference not reflected in execution — the assistant cites AGENTS.md in its response but does not apply its constraints to the actual move
Drift does not always surface as an obvious error. The assistant may still sound plausible and produce useful output. The risk is that unjustified confidence re-enters the session without either party noticing.

​
Re-entering the Contract

When drift is detected — by the operator or by the assistant — the correct recovery is re-entry:
  1. Give the activation instruction again
  2. Reread AGENTS.md in full before continuing
  3. Re-establish the real target and authority
  4. Resume only after the operative frame is active again
Do not attempt to continue the current task while drift is present. Re-entry is the smallest safe unblock move.
For consequential work, ask the assistant to produce the Touch, Ground, State, and Convergence fields at the end of each step, not only at the very end of the session. This keeps the epistemic state inspectable throughout — and makes drift easier to catch before it compounds.

Build docs developers (and LLMs) love

Get started for freeTalk to us