Demo Repository¶
The companion demo repo is available here:
It contains:
- a noisy and contradictory
AGENTS.md - several intentionally messy
SKILL.mdfiles tasks.mdfor task evidenceissues.mdfor recurring feedback themes
Why Use the Demo?¶
The demo shows CodexOpt on a small repo where instruction problems are easy to see:
- duplicate rules
- contradiction
- missing frontmatter
- unnecessary verbosity
Typical Demo Flow¶
The demo is intended to be runnable without needing to invent your own assets first.
It already includes:
- a sample
AGENTS.md - demo skills under
.codex/skills/ tasks.mdfor task evidenceissues.mdfor recurring review themes
1. Clone the demo and enter the repo¶
2. Create or copy config¶
3. Run a baseline benchmark¶
4. Optimize the instruction assets¶
codexopt --config codexopt.yaml optimize agents --engine heuristic --file AGENTS.md
codexopt --config codexopt.yaml optimize skills --engine heuristic --glob ".codex/skills/**/SKILL.md"
5. Review the results¶
codexopt --config codexopt.yaml report --output codexopt-report.md
sed -n '1,120p' codexopt-report.md
Optional: try GEPA¶
export OPENAI_API_KEY="YOUR_KEY"
codexopt --config codexopt.yaml optimize agents \
--engine gepa \
--reflection-model openai/gpt-5-mini \
--max-metric-calls 20 \
--file AGENTS.md
Cross-reference¶
If you are reading the demo first, the main project lives here: