Handling Outputs
Skills produce two kinds of output: text in the conversation (reviews, memos, recommendations) and files (decks, documents, scripts, images). Where files land depends on the environment.
By environment
Claude Code / Codex CLI (terminal agents)
Files are written straight to your working directory (or the path you specify). Best practices:
- Tell the skill where to write: "save the deck to
./outputs/qbr-deck.pptx" - Keep generated artifacts out of git unless they're deliverables; add an
outputs/entry to.gitignorefor scratch work - For decks/documents, open the file locally to QA before sharing; skills like
apd-pptx-designinclude a visual QA step, but human review is still expected
claude.ai / ChatGPT (chat products)
Files appear as downloadable attachments in the conversation. Download them; nothing persists in a shared location automatically. For anything client-facing, move the file into the proper SharePoint/project location; chat threads are not a system of record.
Snowflake Cortex Code / Databricks
Outputs land in the workspace filesystem or stage/volume you point them at. Follow the client project's data handling conventions, especially for anything derived from client data.
Output hygiene rules
- Review before you ship. Skill output is a first draft with standards baked in, not a finished deliverable. The mirrored
rubric.mdfor each skill (undertests/) tells you what "good" looks like; use it as your own QA checklist. - Client data stays in client environments. Don't paste client data into APD-account chat products, and don't move generated artifacts containing client data outside approved locations.
- Name and file outputs like any other deliverable. Project naming conventions apply to AI-generated files too.
- Iterate in the same session when possible. Agents keep context; "tighten slide 4 and fix the chart colors" beats regenerating from scratch.
- Feed learnings back. If a skill's output consistently needs the same manual fix, that's a skill bug; open a Bug or Feedback issue or PR the fix.