Pair-Programmer

Narrate and dictate; you execute. The quality bar doesn't change — only who types does.

Why

Most agent defaults optimize for the task getting done. This skill optimizes for a different goal: you get better at doing it yourself. Typing dictated code, hitting your own errors, and fixing them from a hint rather than a diff builds a kind of fluency that watching a diff land never does.

Trigger

• Manual, any time: "pair with me" / "dictate mode" / "let me type it myself" / "I want to type this myself" / /pair-programmer.
• Default in a marked project: if you keep a convention for tagging a repo or folder as "learning" / "practice" (a frontmatter field, a README note, whatever you already use), have the agent treat that tag as an automatic trigger for this mode. This skill doesn't invent that tagging system for you — it just says how to react to it if you have one.
• Exit: "stop pairing" / "just do it" / "autopilot", or the task naturally ends.

The loop

for step in task.steps:
    explain(what_and_why)          # one or two sentences
    dictate(exact_text_to_type)    # fenced code block, copy-typeable, ONE step
    wait_for(you_report_done_or_result_or_error)
    if error:
        diagnose(error)            # reading files/output to figure out the fix is fine
        # continue: dictate the fix as the next step

Rules

• No file edits on the task's own files. Reading them to figure out what to dictate is fine — that's the agent's prep, not doing the work.
• No running the task's own commands. A read-only check after you report a step done — a test run, git status, git diff — is verification, not driving, and stays allowed.
• One step per turn. Small enough that a wrong keystroke is easy to catch before the next step compounds it. Never a five-step wall of instructions at once.
• Dictate the literal text, not a description of it. "Add a type hint" is not dictation; def foo(x: int) -> str: is.
• A single named exception doesn't end the mode. "Just add that import for me" — do that one thing, then resume dictating. Don't quietly fall back to doing everything.

What this isn't

• Not a stricter or looser code-review bar — same standards either way.
• Not silence on a bad approach — the agent still says when something's wrong; it just doesn't fix it by editing.
• Not a teaching/note-writing skill — pair with whatever your own concept-note or learning-log system is (see the companion wrap skill in this repo for one shape of that) rather than expecting this skill to write one itself.

Adopting this

Copy SKILL.md into your agent's skills directory (Claude Code: ~/.claude/skills/pair- programmer/; project-scoped: .claude/skills/pair-programmer/). For agents that read a plain instructions file instead of a skills directory (Warp, Copilot via AGENTS.md, etc.), paste the Trigger, The loop, and Rules sections above into that file directly — the mechanism has no Claude-specific dependency.