Matching Codex’s Workflow Stack in Claude Code

OpenAI Codex markets four workflow features: Agents.md, custom Skills, Record & Replay, and Automations. Claude Code already covers most of them. This page maps them one-to-one, then shows how to build the two pieces Claude doesn’t ship out of the box.

The map

Codex feature	Claude Code equivalent	Ships by default?
Agents.md (project profile, workflow rules)	`CLAUDE.md` (global + project) + `AGENTS.md` (also read) + memory files	Yes
Custom Skills (`@`-invoke)	Skills (`/name`), authored via `/skillify`, `/autoskill`, skill manager	Yes
Plan Mode	Native plan mode	Yes
Fan-out subagents	Subagents + git worktrees	Yes
Automations (scheduled runs)	`/loop`, `/goal`, plus OS scheduler (launchd/cron)	Partly — no registry view
Record & Replay (demo → skill)	`/skillify` reads the conversation, not your actions	No — build it

The two rows that aren’t a clean match are the build targets below.

Gap 1: Record & Replay (demo → skill)

Codex watches you perform a task and synthesizes a skill. Claude’s /skillify extracts a skill from the conversation transcript — useful, but it never saw the commands you ran in your own terminal or the clicks you made in a browser. Close that with two capture layers feeding one synthesizer skill.

CLI layer — record a terminal session

script (present on macOS/Linux) records a session to a transcript. Drop this in ~/.claude/scripts/demo-record.sh:

#!/usr/bin/env bash
# demo-record.sh — record a terminal session for later skill synthesis.
set -euo pipefail
slug="$(printf '%s' "${1:-demo}" | tr -cs 'A-Za-z0-9_-' '-' | sed 's/^-*//; s/-*$//')"
out="$HOME/.claude/demos/${slug:-demo}-$(date +%Y%m%d-%H%M%S).log"
mkdir -p "$(dirname "$out")"
echo "🎬 Recording → $out   (run your task, then type 'exit')"
script -q "$out" "${SHELL:-/bin/bash}" -l || true   # BSD: script -q FILE CMD; GNU: script -q -c CMD FILE
echo "✅ Saved: $out"

Then a synthesizer skill (~/.claude/skills/demo2skill/SKILL.md) reads the newest transcript, strips ANSI/control noise, extracts the command sequence, asks one or two questions (what varies between runs? what counts as success?), and writes a new SKILL.md. Reuse the conventions in Skills Ecosystem and your skill-authoring checklist rather than inventing a format.

Cleaning the transcript before synthesis (handles both BSD and GNU script noise):

LC_ALL=C sed -E \
  -e 's/\x1b\][0-9];[^\x07]*\x07//g' \
  -e 's/\x1b\[[0-9;?]*[ -/]*[@-~]//g' \
  -e 's/\x1b[@-_]//g' -e 's/\r//g' "$transcript" \
| tr -d '\000-\010\013\014\016-\037' \
| grep -vE '^Script (started|done)'

Browser layer — Chrome DevTools Recorder

Honest constraint. A browser-automation MCP (e.g. chrome-devtools) drives the browser — it cannot passively log a human’s clicks. The clean capture bridge is Chrome’s built-in Recorder panel: DevTools → Recorder → record your flow → Export as JSON.

Feed that JSON to the same synthesizer skill. Each step has a type (navigate / click / change / waitForElement) plus selectors; translate the steps into a readable procedure, and optionally emit a Playwright or MCP playback block so the skill can re-run the flow later.

Why this beats `/skillify` alone

/skillify is great when the work happened in the Claude session. Record & Replay captures work you did outside it — a manual deploy in your shell, a multi-step web console task — and turns the demonstration itself into the skill.

Gap 2: Automations registry

/loop and /goal handle in-session recurring work; the OS scheduler (launchd on macOS, cron on Linux) handles background jobs like a morning digest or a weekly cost audit. What’s missing is a single answer to “what’s scheduled, how often, and when did it last run?” — that state is scattered across plist/cron files and logs.

A small lister gives you the registry view. For launchd:

#!/usr/bin/env bash
# automations-list.sh — one-screen view of scheduled Claude jobs.
set -euo pipefail
la="$HOME/Library/LaunchAgents"; logs="$HOME/.claude/logs"
printf '%-24s %-20s %-18s %s\n' LABEL CADENCE "LAST RUN" STATUS
shopt -s nullglob
for p in "$la"/com.*.claude-*.plist; do
  label="$(/usr/libexec/PlistBuddy -c 'Print :Label' "$p" 2>/dev/null || basename "$p" .plist)"
  if iv="$(/usr/libexec/PlistBuddy -c 'Print :StartInterval' "$p" 2>/dev/null)"; then
    cad="every ${iv}s"
  else cad="calendar"; fi
  log="$logs/${label}.log"
  last="never"; st="-"
  [ -f "$log" ] && { last="$(date -r "$log" '+%Y-%m-%d %H:%M')"; [ -s "$log" ] && st=ok || st=empty; }
  printf '%-24s %-20s %-18s %s\n' "${label#com.}" "$cad" "$last" "$st"
done

Wrap it in an /automations skill with list / inspect <label> (tail the log) / add (scaffold a new plist from a template, then launchctl load after confirming with you — never auto-load a recurring autonomous run). On Linux, swap the launchd scan for crontab -l parsing.

Honest limits

No live capture of GUI clicks. Browser flows go through Chrome’s Recorder export, not passive logging.
Recording needs a real terminal. The record step runs in your shell, not inside an agent tool call — the agent hands you the command to run.
Scheduling is OS-level. There’s no hosted “Automations tab”; you own the launchd/cron entries (which also means full shell power and no vendor lock-in).