Codex CLI 0.164.0, released on October 23, 2026, adds ten numbered /presets slots for reusable prompts and slash commands. That is useful for a short saved instruction. A workflow that also needs a versioned glossary benefits from a skill with reference files instead. The new preset feature does not mean a glossary file is automatically packaged or read; check the actual skill bundle.
A skill does not need to put every reference paragraph inside SKILL.md. In ChatGPT and Codex, skills can include reference files, and instructions are loaded when the skill is selected. A file’s presence alone does not prove the assistant used it.
Give each file one job
Keep the main instructions focused on when to use the skill and how to finish the task. Put a long approved glossary in a named reference file. State when to read it, its version and what to do when a term is absent. Use paths that resolve within the actual skill package.
For example, a product-description skill can point to references/glossary.md and require three controlled terms to remain unchanged. Add one unfamiliar term to the test input: the expected result should flag it instead of inventing an approved translation.
Check the loaded evidence
Inspect the output terms against the reference and check the environment’s available evidence of file access. Progressive loading saves no effort if instructions never direct the assistant to the needed file.
For a small test glossary, list Wi-Fi, USB-C and SSD as approved spellings. Leave sample-only-term out. Ask for a short description containing all four: keep the first three spellings and flag the fourth for review. Record which glossary version was supplied so a later change can be checked.