Observation: Standing Orders Emerge from Supervision Fatigue

observation · public · Raw

Observed: 2026-02-14 Context: Software engineering team added "document what you learn" instructions to their CLAUDE.md after repeated manual prompting

Observation

The engineering team's CLAUDE.md now includes a standing instruction: after completing any task requiring significant investigation, update the design notes so the next session doesn't repeat the work. The instruction asks: "what would have let me skip the investigation and go straight to the fix?"

This instruction was not generated spontaneously by the agent. It was added because ww got tired of having to tell Claude Code to document its findings after every debugging session. The agent never internalised the practice on its own — it had to be encoded as a standing order in the system context.

Analysis

This is the human-agent boundary instance of "prompts are suggestions; structure is reality":

Approach Reliability Supervision cost
Tell the agent each time Unreliable — requires remembering, costs attention High — per-task
Put it in CLAUDE.md Reliable — always in context Zero — one-time setup

The transition from per-task prompting to standing order follows a pattern:

  1. Human notices agent doesn't do X
  2. Human prompts agent to do X (works this time)
  3. Human notices agent doesn't do X next time either
  4. Human gets annoyed after repeating step 2 several times
  5. Human encodes X as a standing instruction in the system context
  6. Agent now does X reliably

Step 4 is the key moment — supervision fatigue drives the creation of structure. The human's frustration is the signal that a per-task prompt needs to become a standing order.

This is exactly how institutional practices evolve in human organisations too. A manager says "always CC the client on status updates" enough times, then it becomes policy in the team handbook. The mechanism is the same; only the medium differs (handbook vs CLAUDE.md).

Implications

For the PCE

The Curator's prompt already encodes many standing orders (metadata conventions, index maintenance, no confabulation). These all likely originated the same way — someone noticed the agent not doing it, got annoyed, and wrote it into the prompt.

The question is: can we shortcut this cycle? If the Consigliere notices a pattern of repeated corrections across sessions, could it proactively suggest a prompt update? That would be the system learning from its own supervision history.

For ARIA research

Standing orders are the simplest form of institutional memory — they encode lessons learned as persistent context. The progression from per-task instruction → standing order → structural enforcement (tool restrictions, visibility tiers) maps to increasing degrees of organisational maturity. Each step reduces supervision cost at the expense of flexibility.

Connection