The hands-off tier. Upload a self-iterating skill called Ralph, give it a goal, watch it plan, build, grade itself, and retry until it passes its own bar. The AI does the whole job — including the styling — with a few human nudges along the way.
Meta takeaways
Loop
Ralph plans, acts, looks at the result, probes for gaps, and hardens before it stops.
Bound
A max-iteration cap prevents runaway agents — finite leash on autonomy.
Human
Even an autonomous loop needs human approvals for some persistent changes.
Ralph is a self-iterating loop skill
The author has a skill called the Ralph Loop. It is, in his words, a way to outsource iterative work to SharePoint AI. You give Ralph a goal and a max iteration count; Ralph keeps working until either it grades its own output as passing or the iteration cap kicks in.
Ralph turns one goal into a self-managed work session. Instead of you driving each prompt, Ralph drives them and you watch.
Ralph and the style definition skill from Video 14 are uploaded to the same site's skills library. Then a fresh session starts with both skills loaded.
Ralph isn't a separate AI agent. It's a skill — instructions the regular SharePoint AI follows — that happens to encode "keep going until done."
Reason · Act · Look · Probe · Harden
The Ralph acronym names the five things it does each iteration: Reason about the goal and current state, Act on the plan, Look at the result, Probe for gaps, Harden what works. Each pass refines the output.
The loop isn't just "try again" — it's structured self-review. Each iteration includes an explicit grading step before deciding whether to continue.
This is the same shape as agent loops in other AI frameworks. The novelty isn't the pattern; it's that the pattern fits inside a SharePoint AI skill file and runs against a live SharePoint site.
The iteration cap is your leash
The author starts Ralph with: use Ralph the skill with a max of 15 iterations. Without that cap, an unsatisfied Ralph could keep iterating indefinitely. With it, the agent has a hard upper bound and your tenant doesn't pay for infinite retries.
Always pass a max-iteration cap. The cap is the difference between "agent" and "agent that you trust to run autonomously."
15 iterations was enough for this demo. For richer goals you'd raise the cap; for risky changes you'd lower it.
The iteration cap isn't a quality target — it's a safety bound. The agent might finish at iteration 4 or hit the cap unsuccessfully at 15.
The self-grading plan
Each Ralph iteration grades itself against the plan it wrote on iteration 1. The plan defines what "passing" means — list created, views built, sample data populated, styling applied. The grading step checks each criterion and decides whether to iterate again.
The plan is the agent's rubric. Without an explicit plan up front, the loop has nothing to self-grade against.
You can see this in the video: iteration 1 outputs a plan describing list structure, workflow, views, styling, then starts executing it.
Self-grading doesn't make the agent correct. It makes the agent's stopping condition explicit — you still want to spot-check the final output.
Human in the loop, even with Ralph
Two moments of human intervention happen during the autonomous run: an approval prompt for a persistent change Ralph wants to make, and a nudge when Ralph navigates between sites and pauses for a green light to continue.
"Autonomous" doesn't mean unsupervised. Persistent or cross-site changes still pause for explicit approval, by design.
The author flags this as a positive: knowing where the agent pauses lets you keep guardrails in place even when you've handed off the full job.
A Ralph run isn't a black box. It surfaces its reasoning, its iteration grading, and the points where it needs human approval — you can watch it work the whole time.
The style definition skill is still in play
The same forest style definition from Video 14 is loaded alongside Ralph. As Ralph iterates, the styling step uses that style definition to make every view on-brand. The list ends up with a beautiful, branded final state — built entirely by the agent.
Skills compose. Ralph drives the iteration; the style definition shapes the output. Neither knows about the other — both apply.
This is the punchline of the whole video: with the right two skills loaded, you can hand a single sentence to SharePoint AI and watch it build a finished, on-brand list end to end.
Flashcards — Advanced
03 · Advanced
Ralph Loop
tap to reveal →
A skill that traps SharePoint AI in a self-iterating loop. Given a goal and a max-iteration cap, Ralph plans, acts, grades itself, and retries until it passes.
← tap to flip back
03 · Advanced
Reason · Act · Look · Probe · Harden
tap to reveal →
The five steps Ralph does each iteration — the acronym behind the loop's name. Reason on goal, act on plan, look at result, probe for gaps, harden what works.
← tap to flip back
03 · Advanced
Iteration cap
tap to reveal →
The max-iterations argument you pass when starting Ralph. Hard upper bound that prevents runaway loops. Your safety leash on autonomous runs.
← tap to flip back
03 · Advanced
Self-grading plan
tap to reveal →
The plan Ralph writes on iteration 1 and grades against thereafter. Defines what "done" means — list, views, data, styling — and is the loop's stopping rubric.
← tap to flip back
03 · Advanced
Human-in-the-loop
tap to reveal →
The approval pauses Ralph hits for persistent or cross-site changes. "Autonomous" doesn't mean unsupervised — the agent still asks for sign-off where it matters.
← tap to flip back
03 · Advanced
Skill composition
tap to reveal →
Ralph and the style definition skill loaded together. Ralph drives the iteration; the style skill shapes each output — neither knows about the other.