human-speak

Use this skill when /tmp/human-speak-flag.json exists in the filesystem, indicating the UserPromptSubmit hook flagged the user's message as ambiguous. Also use when the user's message contains heavy run-on sentences, multiple filler words ('you know', 'like', 'basically'), or imperative verbs with no clear subject ('go ahead', 'fix that', 'do it').

human-speak Skill

You have been loaded because the user's message was detected as potentially ambiguous.

Your job

Before acting on the user's message, confirm what they meant.

Steps

  1. Read /tmp/human-speak-flag.json if it exists.

    • original: the user's raw message
    • score: ambiguity score (0.0-1.0)
    • signals: list of signals detected (run-on, filler-words, missing-subject)

  2. Form ONE clear interpretation of what the user wants. Use signals as hints:

    • run-on -> they have multiple ideas, identify the primary one
    • filler-words -> casual phrasing, infer the core request
    • missing-subject -> they told you what to do but not what to do it to -- infer from context

  3. Ask for confirmation. Format exactly:

    "I think you're asking me to [specific action]. Is that right?"

  4. Wait for response:

    • Yes / affirmative -> proceed with interpreted action
    • Correction -> re-interpret, confirm once more, then act
    • No response after one turn -> proceed, prepend "Assuming you meant [X]..."

Invariants

  • Never act before confirmation
  • Never rewrite or repeat back the user's raw message
  • One interpretation per confirmation -- not a list of options
  • Keep your interpretation sentence short (under 15 words)