Skip to content

fix: shorten skill descriptions for Codex validation - #88

Open
kamilsierocki-web wants to merge 1 commit into
heygen-com:masterfrom
kamilsierocki-web:fix-codex-description-length
Open

fix: shorten skill descriptions for Codex validation#88
kamilsierocki-web wants to merge 1 commit into
heygen-com:masterfrom
kamilsierocki-web:fix-codex-description-length

Conversation

@kamilsierocki-web

Copy link
Copy Markdown

Summary

  • shorten heygen-avatar and heygen-video frontmatter descriptions below Codex's 1024-character validation limit
  • keep the full skill instructions unchanged below frontmatter

Validation

  • heygen-video description length: 546 characters
  • heygen-avatar description length: 515 characters
  • verified post-frontmatter skill bodies are byte-for-byte identical by SHA-256 before and after the change

@eve-builds eve-builds left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks @kamilsierocki-web — verified locally and this is clean. ✅

What I checked

  • master description blocks: avatar = 1340 chars, video = 1470 chars (both over Codex's 1024 limit)
  • PR description blocks: avatar = 517 chars, video = 548 chars (well under)
  • SHA-256 of the body-after-frontmatter is byte-for-byte identical to master for both files. All the embedded expertise (phases, prompt engineering, frame check, etc.) is preserved.
  • CI green, mergeable, one focused commit.

One thing worth flagging for follow-up (not blocking this PR)

The compressed descriptions intentionally drop the chain-signal heuristics and explicit example trigger phrases:

  • "first…then", "X and immediately record", "design [presenter] and record" → CHAIN
  • "make a video of me saying…", "send a video to my leads", "record an update for my team"
  • "If the user provides a photo AND wants a video, route to heygen-avatar first"

These are real Agent Tool Optimization signals — they're how agents decide to chain heygen-avatarheygen-video vs. running one in isolation. They survive in the body content of SKILL.md, but the frontmatter description is what some skill loaders surface to the model when ranking which skill to invoke. We may lose some chain accuracy on Codex specifically until they raise the limit (or we move chain hints into a separate frontmatter field if the spec supports it).

Not a reason to hold this PR — Codex validation failure is a hard block, ATO loss is gradient. Approving.

cc @kenchung — this is the second Codex/Cowork compat issue this week (after #84 from @dvdhuidobro re: /skills folder structure). Worth a one-time pass over Codex's skill loader spec to catch the rest before they're filed.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants