codex-discusslisted
Install: claude install-skill y-a-v-a/skills
You drive a structured, turn-based discussion with OpenAI Codex to converge on a
plan **before** implementation. A second model that argues the other side is
worth more than a second model that agrees with you — so the roles are
deliberately adversarial.
## Topic
$ARGUMENTS
## Why two models
You review your own work with a positive bias and review others' work with a
critical one. This skill exploits that: **you propose, Codex challenges.** Codex
is instructed to find the weakest part of each proposal, not to validate it.
Agreement only counts once the objections have actually been answered.
## When to use this (and when not to)
Turn-based discussion is **not the default** — it costs several minutes per
Codex turn plus the upkeep of a shared artifact. For most design questions a
one-shot critique (a single `codex exec` over a brief, or `/codex-review`)
delivers most of the value at a fraction of the cost. Reach for `codex-discuss`
only when the decision has **durable architecture impact, irreversible
migration cost, or a live Blocker/Major disagreement** a single round can't
settle. Otherwise prefer the one-shot and stop there.
## The shared file
One markdown file is the entire conversation. **You are the only writer** —
Codex runs read-only, reads the file from disk, and returns its turn as text,
which you append. This keeps a single source of truth and prevents write races.
Create it in the invoking repo so the reasoning trail can be committed (or
ignored) as the u