ai-pairing-playbook

The ai-pairing-playbook Claude Code skill — a reciprocal prompting reference for sustained AI pair-programming. Six small shifts the user adopts to make the assistant more effective, reducing drift and improving auto-mode safety.

Designed to pair with dual-log-memory (memory architecture). Together they form the communication and memory layer for long-running AI pair-programming.

The six shifts

1. Pre-name the active guidelines before substantive work.
2. Three flavors of "proceed" — disambiguate authorization for state-mutating steps.
3. Name auto-mode boundaries — narrate where auto ends.
4. Stage multi-part thoughts — type the whole thought in one send, or signal "hold."
5. Make visual reports diagnostic — include what doesn't happen at the end of an observed sequence.
6. Prove the trade-off on a copy — on a high-blast-radius request, test it in isolation before complying; neither capitulate nor stonewall.

Plus a bonus: texture beats brevity for corrections — the why behind a pushback multiplies the lesson's reach.

See ai-pairing-playbook/SKILL.md for the full breakdown, each shift with examples, and the trust-but-verify cross-reference.

Install

macOS / Linux

git clone https://github.com/gmrmk/ai-pairing-playbook.git
cp -r ai-pairing-playbook/ai-pairing-playbook ~/.claude/skills/

Windows (PowerShell)

git clone https://github.com/gmrmk/ai-pairing-playbook.git
Copy-Item -Path .\ai-pairing-playbook\ai-pairing-playbook -Destination "$env:USERPROFILE\.claude\skills\" -Recurse

(The doubled ai-pairing-playbook is intentional — the outer is the cloned repo, the inner is the skill subdirectory that goes into ~/.claude/skills/.)

Usage

Once installed, the skill appears in Claude Code's available-skills list. Invoke via:

Skill ai-pairing-playbook

Adopting in a new project

Copy ai-pairing-playbook/references/PARTNER-NOTES-template.md into your project as docs/working-with-claude.md (or wherever you keep reference docs). Adapt the examples to your project's tech stack and domain. Update it as new mechanics surface — this is a working document, not a fixed ruleset.

The 6 shifts are stable across backends; the location of the partner-notes doc varies. See SKILL.md § Backend portability for filesystem, note-taking-app, and memory-store placement guidance.

Trust-but-verify always

Prompt-mechanics advice is itself a class of claim about how collaboration works. Every recommendation in this skill routes through a four-step funnel (state claim → name second signal → verify → act). Hedged or generic prompt advice is itself the failure mode this skill exists to prevent — session-grounded specifics ("here's the moment X would have prevented") beat generic best-practices every time.

Companion skill

dual-log-memory — A memory architecture pairing a fix log with an insight log. Symmetric trust-but-verify discipline plus sunset rules for self-pruning. Together with ai-pairing-playbook it forms the memory and communication layer for sustained collaboration.

→ https://github.com/gmrmk/dual-log-memory

License

MIT — see LICENSE.

Contributing

Issues and PRs welcome. If you adopt the partner-notes pattern in your own project and surface a new prompt-shift worth feeding back, open an issue describing the moment that prompted the addition. The skill grows by its own discipline — driven by concrete, session-grounded incidents rather than abstract best-practice advice.