Understand and design

Understand before changing

Skill	Use it to
in-your-own-words	Have the agent restate your goals and the problem in its own words.
how	Trace behavior.
why	Investigate motivation from source records, retaining uncertainty.
teach	Combine findings into an explanation.
recall	Reconstruct only scoped history.
session pickup	Resume a specific prior session.

Claude Code

/how trace argument parsing; read-only
/why does the retry path back off twice? cite the commits

Codex

$how trace argument parsing; read-only
$why does the retry path back off twice? cite the commits

Unverified history

Missing verified transcripts get a labeled digest, not a claim of historical coverage.

Sharpen a prompt before sending it

Use sharpen to identify the expert approach, original source material, specific methods and tools, and an independent check, then get a complete rewritten prompt.

Claude Code

/sharpen write a migration guide for the v2 config format; leave the rollout strategy open

Codex

$sharpen write a migration guide for the v2 config format; leave the rollout strategy open

Sharpen preserves your goal, constraints, and open choices. It stops after the rewrite without doing the task. Review the prompt, then send it separately when ready, optionally with z-mode.

Design and review

Skill	Use it to
architect	Design from caller usage and types first.
arena	Settle contested shapes with independent candidates and synthesis.
swarm	Partition coverage across agents.
interrogate	Return adversarial findings that a rerun proves, without edits.

Claude Code

/architect design the cache invalidation API; design only
/interrogate this diff; findings only
/swarm check these three packages; one owned report per package

Codex

$architect design the cache invalidation API; design only
$interrogate this diff; findings only
$swarm check these three packages; one owned report per package

INFO

A design request stops at a design unless implementation is already requested.

Same-model runs

These skills use bounded native agents and report actual identities. Multiple runs may share one model; cross-provider consensus is not promised.

Resolve open questions with prototypes

Use the Prototype playbook when running a small experiment can settle a layout, timing, or behavior question. Compare useful variants in scratch space and inspect the rendered output or measurements before picking one. Skip the experiment when the direction is already established.

Claude Code

/z-mode prototype two settings layouts; show rendered screenshots for comparison

Codex

$z-mode prototype two settings layouts; show rendered screenshots for comparison

For a shared package or API, sketch its caller-facing README or tutorial first. That gives the design a concrete usage target. Resolve experimentally answerable uncertainties before polishing a plan; review can still catch scope, dependency, and authority problems before code exists.

When complexity warrants a plan or the user asks, settle the design, then use the Multi-phase plan playbook to write coherent phases with acceptance and verification checks. The plan is the deliverable; implementation begins only when requested.