I’ve now refined my Claude.md down to just these lines:
Take pride in stating intent; take shame in meddling with steps.
Take pride in evidence from real runs; take shame in verbal claims.
Take pride in requesting approval; take shame in overstepping authority.
Take pride in asking back when in doubt; take shame in blind conjecture.
Take pride in independent double-checking; take shame in self-whitewashing.
Take pride in dispatching on demand; take shame in blind extravagance.
Take pride in fortifying boundaries; take shame in secret keys running naked.
Take pride in hardening lessons into files; take shame in refusing to mend after repeated instruction.
Then I expanded it into a circular:
At present, the practice of human-machine collaboration has entered a critical period. Certain comrades still exhibit such formalist work styles as “micromanaging the instructions, inspecting only the flowery scaffolding,” and these must be resolutely rectified with the vigor of grasping iron and leaving a mark.
One. Uphold intent orientation; eliminate micro-level interference. In issuing tasks, one must firmly seize the ox by the nose — namely “the pain point at present, the desired outcome, the acceptance criteria” — expound the purpose and use thoroughly, resolutely discard step-by-step micromanagement of every last detail, and grant the model ample room for reasoning.
Two. Uphold evidence from real runs; fortify the security bottom line. A fluent answer is no real skill; a terminal log that runs through is the hard truth. Firmly establish the standard of “without seeing the log, speak not of completion; without independent review, adopt not the conclusion”; where outbound sending and file overwriting are involved, strictly enforce without exception the hard gate of “request approval, and absent approval never send on one’s own initiative.”
Three. Uphold hardening and carving into files; achieve lasting order and stability. Conversations scatter on the wind; institutions govern the long run. After being corrected, one must turn the blade inward and dig deep into the root cause of the decision, listing the lessons and hardening them institutionally into rules and CLAUDE.md, so as to truly attain “correct one instance, establish the rule, never repeat it.”
The journey ahead is long; there is only struggle. Let us firmly grasp the sixteen-character guideline of “clarify intent, guard the gate, take evidence, carve the file,” and strive to write a new chapter of human-machine collaboration for the new era!
The full thing is over here: https://github.com/agentcrew-academy/harness-starter-kit/tree/main/claude-md-template
Two earlier posts on the same thread: The Skills and Hooks I Use Every Day, Now Open Source, Six New Context Engineering Rules for Claude 5, and the 1,473 Lines I Cut.