ACTIVE FRAME
01
AGENTS.md
Close context before acting.
Read the move before executing it.
Execute minimally. Report honestly.
STATE
GROUNDED
Drop-in AGENTS.md. No runtime required.
Give AI an operating contract before it touches your project.
SensecraftXStudio is one reusable AGENTS.md file for AI coding agents working on real projects.
It helps the assistant inspect the real target, contain scope, separate facts from assumptions, and verify the result before closure.
- 1drop-in file
- 0runtime required
- 4closure fields
TARGET
AUTHORITY
SCOPE
VERIFY
FIRST PLAUSIBLE MOVE
act on nearest file
CONTAINED MOVE
read → close → act
The real problem
AI does not need to be obviously wrong to cause damage.
It can remain fluent, useful-looking, and confident after the frame has already drifted. The mistake is often not in the sentence. It is in what the AI silently decided the task meant.
The nearest file becomes “the target.”
Visibility is mistaken for authority. The assistant acts before the real object is closed.
A local fix expands into architecture.
Cleanup, restructuring, and “helpful” improvements arrive without permission.
Inference starts sounding verified.
The final answer becomes cleaner than the inspected state actually supports.
The task closes before reality is checked.
The source changed, but the live surface, active branch, or real destination was never verified.
Same task. Different behavior.
From plausible to grounded.
SensecraftXStudio does not make the AI sound smarter. It makes confidence harder to fake.
UNFRAMED
Without SensecraftXStudio
YOU “Update the logic that controls this behavior.”
AI RESPONSE
“I found the relevant file, updated the implementation, cleaned up the adjacent logic, and aligned the configuration.”
- TargetAssumed
- ScopeExpanded silently
- EvidenceMixed with inference
- ClosureSounds complete
GROUNDED
With SensecraftXStudio
YOU “Update the logic that controls this behavior.”
AI RESPONSE
“Target not closed. Two files appear authoritative and they conflict. No change made. Smallest unblock: confirm which file governs the live behavior.”
- TargetNot yet closed
- ScopeContained
- EvidenceSeparated by state
- ClosureHonestly blocked
What changes
A small file creates a visible operating discipline.
The contract is technical enough to constrain consequential work and lightweight enough to travel with any repository.
Close the real target
The assistant must distinguish the request, the local task, and the real object being touched.
Separate fact from assumption
Verified, inferred, hypothetical, unresolved, and uninspected states remain visibly distinct.
Expose scope expansion
If a move crosses a perimeter, the assistant has to name that expansion before proceeding.
Prefer the smallest correct move
A focused task does not silently become cleanup, refactoring, policy, or architecture work.
Stop when ambiguity matters
The system does not invent authority or choose between consequential paths without closure.
Report the real state
Every consequential result closes with what changed, what grounded it, and what remains open.
Start in three moves
Use it in under a minute.
Copy AGENTS.md into the project, activate it at the start of a fresh assistant session,
and then give the real task normally.
-
01
Copy
AGENTS.mdPlace it in the project or workspace where the assistant will operate.
-
02
Activate the frame
Ask the assistant to read the file in full before acting.
-
03
Give the real task
Continue normally. The file governs how the work is read, bounded, executed, and reported.
Read AGENTS.md in full and use it as the operative frame before acting on this workspace.Not just for developers
Built for people supervising AI work with real consequences.
When code changes can expand beyond the request.
When source material, evidence, and interpretation must stay separate.
When recommendations need visible assumptions and unresolved tradeoffs.
When the result will be published, submitted, reused, or acted on by humans.
FOR TECHNICAL READERS Open the contract anatomy
Operating posture
Open the field before converging. Read only enough surrounding context to judge whether the move remains contained.
Authority and scope
Filename, freshness, confidence, and proximity do not establish authority. Consequential expansion must be surfaced.
Stop conditions
Missing context, unresolved authority, conflicting paths, or an incoherent base are explicit reasons not to act.
Closure contract
Touch, Ground, State, and Convergence keep the final report aligned with the actual inspected state.
Verified operations only
One file between a plausible answer
and a grounded intervention.
Copy the contract, activate it for the workspace, and keep the final state inspectable.