Part II: The Reference Community
49. Stewardship Became a Skill
2 min read
At this point, asking AI for the next step was useful but not enough.
The pattern was becoming repetitive:
human asks what next
down
AI proposes options
down
human chooses
down
AI helps execute
down
new knowledge appears
down
knowledge risks being lost
The work needed a new role.
Not only an assistant.
Not only a writer.
Not only a tool recommender.
It needed a Steward.
The Steward's job was to keep the work moving while returning knowledge to the system.
The Steward should:
- read the current source of truth,
- propose the next smallest useful step,
- offer a few options,
- recommend one,
- wait for the human decision when direction matters,
- execute bounded local work,
- record the decision,
- preserve prompts and assets,
- update source files,
- update workflows and skills,
- and return with the next logical step.
This changed the collaboration model.
The human no longer had to remember to ask:
Where should this be documented?
The Steward should ask that question by default.
The human no longer had to navigate the whole folder structure before every small decision.
The Steward should know where things belong.
The human no longer had to manually decide whether an icon, prompt, color or tool choice mattered.
The Steward should treat meaningful outputs as possible community knowledge.
This became the first project skill:
aifc-reference-steward
The skill was stored in the project source of truth and installed into Codex as a runtime copy.
That distinction mattered.
The installed skill lets the agent work.
The project copy lets the community own the skill.
The principle became:
Skills that shape community capability must live in source of truth, not only in an AI tool.
The Steward was not created as a convenience.
It was created because the reference community had started to experience the exact problem AIFC was designed to solve:
AI accelerated work, but acceleration created new knowledge faster than the human could manually preserve it.
The answer was not to slow down.
The answer was to create a better learning loop.