Derek Wang's agent charter loads the refusal list before the first prompt

Four of the ten rules specify an ordering: read before write, inspect B, C and D before changing A, watch a test fail before fixing it, and read the disk before asserting what a file contains ^[8]. A harness can check every one of those against its own tool-call log ^[1]. The log settles it, whatever the model says about itself.

Two more leave artifacts behind. Scalpel edits show up as diff size, and writing the lesson back shows up as a record that got appended ^[2]. The three-strikes rule says to change the tooling the third time the same kind of failure happens ^[8]. That needs a failure count that outlives the session, and keeping the count is the team's job.

The rule with the best provenance is trust the disk, not memory. Wang says it cost him real time: an agent insisted a file was an old version because that was what the chat memory held, while the file on disk had long since changed ^[9].

The flattery clause is the one the model grades itself on. "A language model is a probability machine that leans toward whatever keeps the conversation comfortable," Wang wrote ^[11]. His answer is a written clause the model runs before it emits anything, ending in stripped polish or "I don't know" ^[12]. He argues a human organization has no equivalent, because the same defect in a human teammate eventually gets called out and in a model never does ^[17]. The model is the only thing checking that self-check. The companion rules around it can be checked from the text: banning "great question!" and "happy to help!" is a string match, and requiring the model to state uncertainty in the first sentence is a position a checker can look for ^[13].

The charter also claims to set a coordinate system. In one of Wang's projects its first sentence is "a general-purpose coding-engineering methodology and multi-agent orchestration system" ^[5]. He says the sentence matters less for its content than for being the first thing the model reads ^[6]. "Get that wrong first, and no amount of clever prompting downstream will save you," he wrote ^[14]. The essay makes that case by argument and by the stale-file incident; it contains no measurement either way ^[18].

For the claim to transfer, several things have to hold in your setup. The file has to be loaded at the start of every session, which is what the contract asks for ^[2]. The boundary list has to name stacks and directions you genuinely refuse. Wang says the refusals are the part that keeps the model off any direction that merely sounds reasonable: "What you refuse to do matters more than what you do," he wrote ^[4]. And the rules have to survive whatever context trimming your harness does between turns. Wang says none of the ten depend on a specific model or framework, and that having an agent read the list before it works is "money in the bank" ^[15]. That last part is untested in the essay, and a team adopting it is writing and reviewing a second document.