Vrataski — Vision & Direction

A distillation of the thinking developed in conversation on 2026-06-13. This document captures a pivot in intent: the decision to unify Yggdrasil, the personal layer, and the execution-partner tools (Odin Codin' and successors) into one cohesive project — an accessibility tool for neurodivergent people — and to make that unified thing what "Vrataski" becomes.

Status — read this first. This is a direction, not a plan and not a commitment. The foundations (Yggdrasil, the personal layer, Odin Codin') already exist and work; the unification into "Vrataski as an accessibility product" is a stated aspiration, not yet designed or built. Much of the product/onboarding material below is explicitly much-later horizon work. The honest near-term reality is that the current foundations still need shoring up (e.g. the current-plan.md redesign). The natural next step, when there are spoons for it, is a proper /brainstorming → /planning pass on the pivot itself. Nothing here is decided by being written down — it's captured thinking, kept so it doesn't evaporate.

1. The pivot, in one breath

What started as a personal toolkit for dogfooding AI — Yggdrasil (the plan/design partner), the personal layer (how Brad works), and execution partners like Odin Codin' (code comprehension) — is being reframed as a single, cohesive accessibility product. The through-line that was always implicitly there becomes the explicit purpose: this is an accessibility tool for people like me — people whose executive function, not their ability, is the barrier between what they can think and what they can get built. That product is Vrataski.

The two things this document has to hold together are a design ethos (human-in-the-loop at all times, never autonomous) and an accessibility purpose (removing the executive-function execution barrier while preserving the person's agency). The central claim of the pivot is that these are not two ideas — they are the same idea seen from two sides. Keeping the human at the center is exactly the property that lets the tool carry an executive-function load without carrying off the person's agency with it.

2. The design ethos — human-in-the-loop, never autonomous

The stance

Most AI tooling today is built to run on its own: hand it a task and it executes the whole thing start to finish, leaving the human to either trust the output or untangle it. Vrataski is the deliberate opposite. It takes the shape of those autonomous systems but puts a person back into every step. The human stays in the loop the whole way through, and the AI never decides or acts on its own. There is no point where it quietly takes the wheel — there is no autonomous loop running that the human is merely supervising.

This is the distinction that matters most, and it's a category difference, not a settings difference. Most tools' "human-in-the-loop" means gating an autonomous loop — approve each step of an agent that is still fundamentally driving. Vrataski means something stronger: there is no autonomous loop to gate. The human is not the exception handler the machine escalates to when it breaks; the human is the clock — the per-step pause is the mechanism, not the failure mode.

The lineage (how it was built)

The architecture is, concretely, a de-autonomized fork of an autonomous AI-development pipeline (the obra/superpowers skill system: brainstorm → plan → execute). Yggdrasil took that pipeline's structure and, at every joint where an agent acted autonomously, substituted a human — recording the rationale each time:

Subagents grinding through tasks unattended → execution is a collaborative walk through the plan, task by task.
Automated, blocking test gates → "the work checks out" is a question the human answers; a machine check is evidence the human weighs at a checkpoint, never a substitute for the human seeing it.
Tasks sized to the smallest agent step → tasks sized to a meaningful human review checkpoint.
Skills that auto-chain into one another → human-pulled seams: each stage recommends the next and the human invokes it (enforced down to the tooling — the planning skill literally cannot auto-invoke itself).
Pressure/compliance language ("you MUST") engineered to coerce a model → a calm house style that marks genuine hard stops sparingly and leaves everything else soft prose; forceful phrasing is kept only as a scoped tool for high-structure execution, never as the default for collaborative work.
Parallel autonomous agents → parallelism is allowed, autonomy is not. Independent agents may run only as read-only advisory critique lenses whose findings route back through the human, who remains the sole decider.

The one-line signature: a de-autonomized, domain-generalized, calmer version of an autonomous coding framework — every departure either removes an assumption (code, rigid test-discipline, agents grinding tasks) or softens a pressure (hard gates, auto-chaining).

The character: it proposes, the human disposes

What the AI is, under this ethos: a thinking partner and a set of hands — never a replacement for the judgment. It proposes; the human disposes. It is deliberately calm — it offers and suggests rather than pushing or pressuring — because a tool that nags or demands is a tool the person comes to resent and avoid. (See §3: for a demand-avoidant brain this isn't politeness, it's the difference between usable and unusable.)

Mechanisms that exist only because a human is in the loop

An autonomous executor never needs these; a human who pauses, forgets, defers, and has variable energy does:

Tracked deferral — "not yet" is a first-class, recorded state, with honest soft triggers ("no spoons for this right now") treated as fully valid. An autonomous system has no "build it later" tier; a human choosing not yet creates it.
Save-and-resume across context loss — every piece of work is resumable from durable disk state, so the person can walk away and come back without losing the thread.
Anti-rot without nagging — accumulating lists and deferrals are surfaced gently and informationally, never as demands; nothing is ever lost (pruned items are archived, not deleted), and nothing pesters.
No-demand-character design — flexibility and anti-rot framed as accessibility, not polish. Rules are flexible defaults the person can take or leave, because demands framed as demands trigger avoidance. Motivation works the same way: it's intrinsic — the pull of a real win on a problem the person actually brought — never streaks, gamification, or nags.

3. Why it is an accessibility tool

This is the purpose the pivot makes explicit. (The first-person external statement of it lives in the companion file vrataski-pitch.md; this section is the underlying argument.)

The barrier is executive function, not ability. For a person who is autistic and has ADHD, the hard part of getting things done is often not the thinking — the problem-solving, the design, the engineering judgment — but the executing: task initiation, working memory, shifting gears, and the sheer mechanical grind (in software: the literal typing and syntax). That gap between can think it and can get it built is executive dysfunction, and it is precisely the part Vrataski takes over. Because the AI stays in the loop and the human stays in charge, it carries the costly execution without taking the work away.

It runs at the pace of whatever energy you have. The apparent paradox — doesn't "in the loop at every step" ask more of a person, not less? — dissolves because the loop scales to available spoons rather than imposing a fixed tax. Spend a lot on a good day and make a lot of progress; spend a little on a low-spoon day and make a little — both are fine, because the work is durable and resumable, so there's no minimum to clear and progress is simply commensurate with what you put in. You work at your own pace. And the decisions it does need are tuned to cost as little as possible: a low-cardinality choice arrives as a single button-press rather than an open-ended prompt, so staying in charge stays cheap. You remain the decider; the price of deciding is filed down to near nothing per step.

It's domain-agnostic, because the wall is in the doing, not the subject. Software engineering is the first-class case, but the same tool helps execute anything where the barrier is the legwork rather than the decisions — planning a trip, a budget, a document. The executive-function barrier isn't specific to code, so neither is the help.

It externalizes memory and reasoning, which is itself an accommodation. The work, its whole history, and — crucially — a running log of why each decision was made live in plain documents and version control rather than in anyone's memory. A person can step away for twenty minutes or a year, return, and be reoriented — and not just to what was decided but to the reasoning behind it, which is the part hardest to reconstruct and most valuable to recover. For a brain that drops threads and bounces between things, that durable record is the difference between resuming and starting over. (This is the "extended mind" idea — external tools as genuine functional extensions of cognition — that the research literature in §4 names directly.)

The interaction itself costs less. Working with the tool doesn't require the masking a person would do alongside another human — the constant low-grade labor of managing how one comes across — and the deliberate softness (offering, never pushing) is exactly what makes it workable for someone who is demand-avoidant, including under self-imposed pressure.

It's strengths-based and self-determined by construction. The person does the reasoning and makes every call; the tool removes a barrier rather than correcting a deficit. This is the accessibility framing that already lives, in seed form, in the personal layer's "design without a demand character" principle (the autonomy/PDA + variable-spoons accommodation). The pivot promotes that principle from how the author works to what the product is for.

4. Research grounding

The framing is not idiosyncratic; a real and growing literature lands in the same place, treating generative AI as assistive scaffolding for executive function rather than a replacement for thought.

The closest single source — a 2025 longitudinal case study (Mittler, Harnessing Generative AI to Overcome Executive Dysfunction in Higher Education, HEAd'25; read in full) — concludes that when AI is used as an "assistive scaffold rather than a replacement for critical thought" (process support, not content generator), it removes executive-function barriers (task initiation / "analysis paralysis," working memory) and thereby enables "a fuller expression of [the person's] capabilities." It frames this through the Extended Mind thesis (Clark & Chalmers) — external tools, including external memory, as functional extensions of cognition — and calls the result "an evolution, not an erosion, of intellectual endeavor," with the human as director and the contribution becoming strategic orchestration.
The broader corpus echoes a recurring line: generative AI as a low-barrier assistive technology that adapts to messy, nonlinear cognition rather than demanding conformity to rigid workflows — which is the no-demand / demand-avoidance point almost verbatim.
A useful nuance that flatters this project's approach: several papers criticize mainstream ADHD/neurodivergence tech for framing neurodivergence as a deficit to correct and for excluding neurodivergent people from the design. Vrataski is built by the neurodivergent person it serves — the strengths-based, self-determined design those critiques say is missing.

(Fidelity note: the Mittler paper was read in full; the rest is from abstracts and search results. Solid enough to establish that the framing has scholarly support; specific statistics should be verified against source before being quoted externally.)

Key sources to pull in full when this goes external-facing: Mittler 2025 (HEAd'25); Executive Dysfunction by Design (ASSETS 2025, ACM SIGACCESS — the prime candidate for library access); Carik et al. 2024, Exploring LLMs Through a Neurodivergent Lens (ACM HCI; on arXiv); the npj Digital Medicine 2024 systematic review on AI for adaptive functioning in neurodevelopmental conditions; and the CHI 2025 study on AI coding assistants for visually impaired developers.

5. The architecture being unified

The pivot doesn't discard the existing structure — it gives the existing layered architecture a product purpose. The pieces and how they relate:

The layered model (composition, not forking). Personal layer (how the human works) → Yggdrasil (the plan/design partner and shared substrate primitives) → execution partners / workflows (domain toolkits like Odin Codin') → project-specific layers. Each layer composes on the one above via symlink aggregation rather than by forking or subclassing — "dependency injection, not inheritance." This is the structural alternative to maintaining one giant inherited skill tree.
Yggdrasil is the front of the pipe: brainstorming (vague idea → validated design) and the flagship planning (design → executable, checkpoint-able plan), both human-in-the-loop and domain-agnostic. The plan is the hand-off artifact between design and execution.
Odin Codin' is the first execution partner — collaborative, incremental, human-in-the-loop code comprehension (propose → confirm → commit), producing durable, AI-facing knowledge artifacts. It applies the identical human-replaces-autonomy inversion to codebase understanding. Currently a skeleton; content lands in later phases.
The recursive plan mechanism is the unifying frame: one artifact and one act repeated at every layer — a markdown plan walked through with an agent, human-in-the-loop, git-backed. Domain knowledge lives in the toolkits; Yggdrasil stays domain-agnostic.
The session bookends (/good-morning, /save-progress) provide the save-and-resume outer loop: orient from a durable checkpoint at the start, pin state at the end, with git-worktree isolation so every session is an atomic, resumable unit.
The hygiene / anti-rot system (bookmarks, backburner, archive, a machine-maintained ledger) keeps the accumulating state lean and surfaces what's due without nagging.
Memory and durable docs — per-fact memory files, a durable-docs manifest, and the decision logs — externalize knowledge so a fresh session is genuinely caught up after a context wipe. (Notably, this layout independently converged on the personal-knowledge- management pattern of "one fact, one file, path as key, frontmatter as schema.")
Markdown-everything is the load-bearing substrate choice: because every artifact the agent reads and writes is plain markdown in git, the system's state and history are fully transparent — no hidden working memory, no silent profile, nothing the user can't open and read. (The underlying model is of course still a black box; the stance is to treat it as a relatively blank slate — raw capability the transparent system is built on top of, with everything we build kept in the open.) That's what makes the system's transparency structural rather than a feature.

The unification task (much-later, to be designed via /brainstorming) is to turn this from "Brad's personal toolkit composed of separate repos" into "one coherent product a person like Brad could adopt." That is the work the pivot names but does not yet do.

6. Product ideas (extracted to its own doc)

The concrete product / onboarding ideas — onboarding that embodies the ethos, progressive disclosure of understanding, the calibration prompt-bank, the hand-written intro, the transparent "what the human learned" user-model, the markdown-everything transparency window (Kingdom.md) — now live in their own file, vrataski-product-ideas.md, so this vision doc stays a tight read. All of it is aspirational, much-later horizon work; see that doc for the detail.

7. Positioning & distinctiveness

The five-part signature (no single piece is novel; the combination had no found replica): de-autonomized + domain-general + accessibility-first + human-pulled seams + semi-rigid-with-calibrated-escape-hatches.
Against the zeitgeist's grain, on purpose. There is a genuine 2026 swing toward "orchestration, not autonomy," so the stance is timely, not contrarian-for-its-own-sake — but Vrataski pushes two steps further than that camp: it removes the autonomy rather than supervising it, and leaves the code domain entirely.
The name. Vrataski is Rita Vrataski from Edge of Tomorrow — the human in the driver's seat of an exosuit, against a film full of autonomous machines; "live, die, repeat" is the save-and-resume iteration loop. The brand is the thesis: the human is always the driver; the AI is the exosuit, never the pilot.
The honest trade-off. The human-in-the-loop pace is affordable precisely because this is personal-scale — it isn't racing to onboard a team or ship a migration. That's a deliberate trade (depth, control, and agency over raw throughput), not a free win, and a product built on this stance should own it rather than hide it.

8. Open questions

Genuinely unresolved as of this writing — named rather than papered over, since this is an idea document, not a settled design:

Who, exactly, is this for? The thesis assumes a person who can do the higher-order thinking and is blocked on execution — a specific profile, not all neurodivergent people. The software case may really be "for neurodivergent software engineers," with the domain-agnostic angle (trips, budgets) being what broadens it to "people blocked on executing anything." Unreconciled, on purpose, for now.
Personalization vs. generalization. The tool's power today is its deep fit to one person's brain. "For people like me" needs either per-user adaptation (the onboarding/calibration path in the product-ideas doc) or a fork-and-fit-it-yourself framework — different products, not yet chosen.

9. Status, risks, and the honest next step

This is a direction, not a plan. Nothing here is built or committed by being written down.
Foundations come first. The existing substrate still needs work before any productization (the current-plan.md redesign is a live example), and most of the product ideas are much-later horizon work.
Personal-first remains the safe default. The tool exists, today, to serve one person well. "If it grows into something for others, fine" has always been the posture; the pivot raises the aspiration without abandoning the personal-first reality. Building and living in it (dogfooding) remains how the design earns its keep.
The honest next step, whenever there are spoons for it, is a proper /brainstorming pass on the pivot itself — treating "unify the layers into an accessibility product" as a design problem to walk through, not a decision to execute. No pressure, no timeline; this document exists so the thinking is durable until then.

Companion artifacts: working/vrataski-pitch.md (the first-person, jargon-free external pitch) and working/vrataski-product-ideas.md (the extracted product / onboarding ideas). Related background captured this session: the human-in-the-loop ↔ obra/superpowers melding analysis (in conversation) and the executive-dysfunction research scan (§4).