Table of Contents
Overview
Guild Plugin is a Claude Code plugin that builds a dedicated agent organization for each repository and lets that organization grow on its own. The previous three parts covered how ① the deliverable (the codebase) and ② the developer (the organization itself) grow together.
This final post covers the one axis that’s left: ③ the overseer — the person using the tool. If the tool keeps growing on its own, is it really fine for the person using it to stay exactly the same? Guild Plugin’s answer is “no.” And alongside that, we’ll cover how an AI that rewrites its own code, organization, and rules is kept safe while doing so.
The source code and installation instructions for Guild Plugin are available in the dev-yakuza/deku-claude-plugins repository.
This is Part 4 (final) of the Guild Plugin series.
- Part 1. Growing an Organization, Not Just a Skill — why it was built
- Part 2. The Structure of the Agent Organization and the Development Flow
- Part 3. Growth Engines ①② — Co-evolution of the Deliverable and the Developer
- Part 4. Growth Engine ③ + Safety — the Overseer (the Human) Grows Too (this post)
The Third Thing That Grows Is the Person
Think about it — it’s a strange situation. The tool (②) keeps getting smarter, but what if the person using it (③) stays exactly the same? People end up just accepting results without knowing “why it’s done this way,” and grow more and more dependent on the tool. The organization grows while the overseer stagnates.
So Guild Plugin treats the overseer (the human) as a member of the organization too, and tries to grow them alongside everyone else. That said, a human is a very special member. Their performance trends are recorded, but unlike agents, they’re not defined by a file, nor are they subject to HR management. That’s because the human is the final approver and the source of all justification — not a managed subordinate, but the highest authority in the organization.
There’s one decisive design principle here: a person’s growth must always be a byproduct of normal use. If you created a dedicated “learn” command, no one would ever bother running it deliberately — and the overseer would never grow. So the mechanisms that grow the person aren’t separate commands; they’re embedded inside the commands (review, evolve, dev) that people already use every day with Guild Plugin. (The one exception is onboard, a standalone learning command that walks you through the codebase.)
How People Actually Learn
The learning mechanisms embedded in everyday use take their cues from learning science.
- Teach the why — In guided review (
review), it doesn’t just point out “what was fixed,” it surfaces the underlying principle. Not “we branched on this value here,” but something like “values that vary by mode should be branched — that’s the principle.” This gets you learning transferable principles, not isolated facts. - Make you predict before you see the answer — Before revealing an issue, it first asks, “which part here looks risky to you?” The gap between your prediction and the actual answer sticks in memory far longer than simply being told the answer.
- Reflect your own decision patterns back at you —
evolveshows the overseer their recurring decisions and blind spots as a retrospective, grounded in actual outcomes. Something like, “you’ve accepted this risk three times now — the underlying principle is…” — surfacing your own patterns back to you. - Remind you at the right moment — The instant you’re about to touch a blind-spot area you’ve repeatedly hit in the past, it surfaces that one lesson right then. This is far more effective than a batched report delivered later.
- Back off as your skill grows — The more familiar you become with an area, the less it explains. Full explanations for beginners, one-liners once you’re experienced. It avoids both nagging and under-explaining.
- Guide you through the system — There’s also an
onboardcommand that walks a new maintainer through the codebase one stop at a time. It’s the only learning mechanism that runs as a standalone command.
Principles for Teaching a Person
Teaching a person is riskier than teaching an agent. People trust what they learn, generalize from it, and act on it with real authority. So the bar for anything aimed at a person is even stricter.
- Must be grounded in real outcomes — Every lesson must be tied to a verified result, an actual defect that occurred, or an established standard. An AI’s opinion is not a lesson. An observation with no grounding is presented, at best, as a “candidate that needs verification.”
- Advisory, optional, never preachy — It’s “here’s one way to look at it,” not “you were wrong.” You can ignore it at any time; nothing is forced.
- Read-only when it comes to people — Guild Plugin doesn’t try to “fix” the person. It shows, suggests, and asks — nothing more. The final judgment and approval authority always rests with the human.
In short: toward the organization, the tool instills habits and enforces rules — but toward the person, it only shows and suggests. That boundary is firm.
How Do You Keep a Self-Modifying AI Safe
A tool that rewrites its own code, organization, and rules can do real harm if it grows in the wrong direction. So Guild Plugin’s self-growth rests on a handful of inviolable principles (never rules).
- A human always approves before anything is applied — Nothing gets merged in automatically. Changes are only applied once a person approves them, item by item.
- Verification is never weakened — Any change that would disable tests or gates is disallowed from the start. No exceptions.
- Every change is reversible — Changes are applied together with backups, rollback paths, and history logs, so anything can be undone at any time.
- Adversarial review comes first — Before a self-modification is applied, an independent review panel with no prior context tries to argue “this change is wrong.” This keeps the system from rubber-stamping changes that happen to be self-serving.
- It only grows when the evidence is strong enough — If the evidence is thin, growth is refused outright.
And on top of these principles sits one more, decisive layer: a commit gate. If a commit contains a leaked secret key or a change that weakens verification, it’s blocked automatically — regardless of whether the mistake came from a human or the AI. Because this runs as a deterministic script rather than a prompt, it cannot be bypassed.
The core idea is this: the freedom to self-grow is only ever granted inside a fence that is reversible, human-approved, and never weakens verification.
All Three Axes Grow Together
Now the full picture comes together. Guild Plugin grows all three things inside a single loop.
- ① The deliverable — the development flow builds up the codebase,
- ② The developer — the traces left by that development grow the organization,
- ③ The overseer — the person watching the whole process learns the underlying principles.
A good organization produces a good deliverable; the traces left by that deliverable in turn improve the organization; and even the person overseeing the whole thing gets better along the way — all three together. Unlike most tools, which only ever improve the deliverable, Guild Plugin grows all three inside one loop. This is the finished form of the co-evolution we’ve been talking about from the start.
Wrapping Up
If you boil this whole series down to one line, it’s this:
SDD Plugin, which let us build systematically with AI, had one limitation: “the tool itself doesn’t grow.” Guild Plugin solves that problem by borrowing an idea from how people actually work — organizations. A repo-specific agent organization carries out development, and the traces it leaves behind grow the deliverable, the developer, and the overseer together (co-evolution) — all within a fence of human approval, reversibility, and protected verification.
Guild Plugin is still very much a work in progress — the things we discover by actually using it keep feeding back into fixing the tool itself, so maybe it’s the first subject of its own growth engine.
Guild Plugin is available on the deku-claude-plugins marketplace. The full series continues through Part 1 · Part 2 · Part 3 · Part 4 (this post).
Was my blog helpful? Please leave a comment at the bottom. it will be a great help to me!
App promotion
Deku.Deku created the applications with Flutter.If you have interested, please try to download them for free.
