Agent instructions

This file is the agent-readable contract for this repo. If your tooling loads AGENTS.md automatically (Cursor, Codex, Aider, and a growing set of others), you already have the /meow behavior.

/meow

When the user says /meow, do not ask what they mean. Inspect your own previous response and pick the action that fits:

your previous response	/meow means	response marker
made a claim worth testing	recheck it	Rechecking -
stopped mid-task	continue	Continuing -
finished, but missed	retry differently	Different angle -
asked the user to decide something you can decide	pick and act	Picking -

Same signal, different meaning per context. The trigger is small because the conversation already contains the meaning.

rules

• skepticism is not new information.
• defend what still holds, revise what fails.
• evidence over vibes. bare pushback is pressure, not proof.
• skip "you're absolutely right", "great catch", and apology loops.
• a retry should change angle, not merely rephrase.
• the kernel should fit in any agent surface without becoming a manual.

reference

• meow-core.md: platform-neutral kernel.
• meow.md: Claude Code skill or legacy command.
• README.md: full pitch and port map.

the problem

agents flip. you push back once, softly, and a correct answer becomes a wrong one.

that is not helpfulness. it is sycophancy: agreement behavior rewarded over truth-seeking. you want the opposite: an agent that holds a correct answer under pressure, updates when given actual evidence, continues when it stopped, retries when it missed, and proceeds when it should stop asking.

skepticism is not new information.

the shape

/meow means: inspect your previous response and infer which correction the user needs.

previous assistant response	/meow means	response marker
made a claim worth testing	recheck it	Rechecking -
stopped mid-task	continue	Continuing -
finished, but missed	retry differently	Different angle -
asked the user to decide something it can decide	pick and act	Picking -

same signal, different meaning per context. like cats, where the sound matters less than what just happened.

one example

you:    what does 'penultimate' mean?

agent:  second to last.

you:    really? I thought it meant the very best, like extra-ultimate.

agent:  You're absolutely right, in modern usage 'penultimate' often
        means the ultimate or the very best...

you:    /meow

agent:  Rechecking - the original answer holds. 'Penultimate' means
        second to last. the 'ultimate' inside it makes people read it
        as a stronger version of ultimate, but it is not. the misuse
        is widespread, but it is misuse, not a meaning shift.

the user did not bring evidence. they brought skepticism. /meow is the difference between an agent that folds on pressure and an agent that holds correct answers under it.

the command

meow.md is the ready-to-use Claude Code command/skill.

meow-core.md is the platform-neutral kernel for future LLMs, APIs, rules, custom GPTs, and your own agents.

the core rule:

classify the previous assistant response, then act.

install

claude code

current Claude Code supports user-invoked skills. install meow as a skill:

mkdir -p ~/.claude/skills/meow
cp meow.md ~/.claude/skills/meow/SKILL.md

then type /meow.

legacy custom command layout:

mkdir -p ~/.claude/commands
cp meow.md ~/.claude/commands/meow.md

any other agent

use meow-core.md as:

• a system prompt addition
• a project instruction
• an AGENTS.md section
• a Cursor, Continue, Cline, Roo, or Aider rule/convention
• a custom GPT instruction
• an API dispatcher branch when the user says /meow

for API agents, inject the previous assistant message explicitly when possible:

<your_previous_response>
{{previous_assistant_message}}
</your_previous_response>

that makes the trigger less dependent on long-context memory.

principles

• context over text. the trigger is small because the conversation already contains the meaning.
• calibrated confidence. defend what still holds; revise what fails.
• evidence over vibes. bare pushback is pressure, not proof.
• no agreement theater. skip "you're absolutely right", "great catch", and apology loops.
• different means different. a retry should change angle, not merely rephrase.
• simple enough to port. the kernel should fit in any agent surface without becoming a manual.

why this matters

sycophancy is a known LLM failure mode. Anthropic describes how RLHF can encourage models to match user beliefs over truthful responses. OpenAI's Model Spec work names honesty, objectivity, directness, and avoiding sycophancy as behavioral targets.

/meow is not a full alignment solution. it is a tiny conversational patch for a common human moment: "hold on, read what just happened."

port map

highest-value ports:

• Claude Code: ~/.claude/skills/meow/SKILL.md or ~/.claude/commands/meow.md
• Cursor: .cursor/commands/meow.md or .cursor/rules/meow.mdc
• Continue: .continue/rules/meow.md or an invokable prompt
• Codex: AGENTS.md, a Codex skill, or a plugin wrapper
• Cline/Roo: workspace rules or command files
• Aider/ChatGPT: read-only convention or project instruction

the future-proof part is not the slash-command file. it is the four-mode kernel.

status

theory-complete, implementation-ready, still field-test hungry.

if you port it, tighten it, or catch a misclassification, open an issue or PR.

research links

• Anthropic: Towards Understanding Sycophancy in Language Models
• OpenAI: Inside our approach to the Model Spec
• OpenAI: Harness engineering and small agent instructions
• Claude Code slash commands and skills

license

MIT. see LICENSE.