Yggdrasil Backburner

Cold storage for Yggdrasil's someday/maybe items — bigger moves filed as ambition, not actionable now. Reviewed on cadence (not loaded into context); promote to an active bookmark or a Target when an item warms up. Entries are dated prose bullets.

  • 2026-05-27 — Agent teams as the next expansion path for Odin Codin' when subagents stop being enough. Requires Claude Code v2.1.32+ with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Higher token cost but true peer-to-peer coordination.
  • 2026-05-27 — Self-hosted web UI / building a Hermes-or-OpenClaw-style harness. Filed as ambition, explicitly not this project. Revisit if Claude Code itself starts feeling constraining.
  • 2026-05-27 — LSB social-license conversation. Required before Phase 10. The plan: approach the LandSandBoat community openly, frame the project as community-internal documentation rather than outside AI scraping, ask before publishing anything. Project's own design (human-in-the-loop, knowledge capture not training) is the best defense against AI-skeptical objections.
  • 2026-05-27 — Public release of Yggdrasil and/or Odin Codin'. Currently both private. Could open-source either or both eventually — Yggdrasil is more naturally shareable since it contains no personal preferences. Decide after they've stabilized. Proposed name (2026-06-11): Vrataski.ai for Yggdrasil — see the naming bookmark in bookmarks.md for the Edge of Tomorrow rationale (human-driven exosuit "jacket" vs. autonomous robots; "live, die, repeat" iteration).
  • 2026-05-27 — creating-workflow-skills — Yggdrasil's first real "create things" skill (the meta-skill that creates other skills). Handles the full procedure of adding a new skill to any workflow (mkdir, write SKILL.md, run skill-wiring to wire it in, verify discovery), and bakes in the disable-model-invocation / allowed-tools design questions so future skills answer "safe to auto-fire?" / "which tools to pre-approve?" deliberately rather than by default. The natural long-term home for the "re-run relink after adding/renaming/removing a skill" habit (a CLAUDE.md note carries it meanwhile). Lives in Yggdrasil but parameterized to target any workflow. Good fresh-session-and-build candidate — it'll test the wiring of whatever it just created. Prior art to diff before greenfielding (2026-05-29): superpowers' writing-skills skill (github.com/obra/superpowers) and the Anthropic skill-creator plugin — compare both before writing our own. (Consolidated from active bookmarks 2026-06-09.)
  • 2026-05-27 — Build a /skill-eval or /skill-benchmark command for Yggdrasil once enough skills exist that drift becomes a concern. Pattern exists in Anthropic's skill-creator (run_eval.py, aggregate_benchmark.py) and community skill-optimizer repo. (Backburnered 2026-06-09 — gated on "enough skills that drift becomes a concern.")
  • 2026-06-02 — Use Yggdrasil to design a personal journaling agent tailored to Brad and his autism — possibly with functionality to summarize entries for his therapist. Strong fit: a non-code design instance (like the learning project), a high-value dogfooding case (a tool built around how Brad actually thinks/processes), and adjacent to the learning-project / AI-learning bookmark in spirit. Privacy and sensitivity are first-class design constraints from the start (therapist-summary export especially). Just capturing — design later via /brainstorming. (Backburnered 2026-06-09 — coldest of the cold, per Brad.)
  • 2026-06-08 — Build a proper FFXI client-RE project in Yggdrasil. A dedicated workflow/toolkit (its own repo, or a Yggdrasil-composed layer) for Final Fantasy XI client reverse-engineering — DAT/texture/format decoders plus validated findings — developed while puppeting the XI-RE-VibeTesting target. Why: the "accumulate here, port later" arrangement (Brad, 2026-06-08) — build and validate knowledge in our own well-structured project, then push it into the foreign target repo in one consolidated pass rather than piecemeal read-only-boundary-crossing edits into a target that will eventually be superseded. Seeds already in the puppet workspace (.puppet/C-Projects-XI-RE-VibeTesting/): research/scripts/graphic_decode.py (general FFXI image decoder — 8bpp-paletted / 32bpp BGRA / DXT1/3, any dimensions), and solved findings in research/ + PUPPET.md (status flag field, PS2 fixed-point alpha, BMP2/Graphic + menu-container formats; the moji menu-container block-order/swizzle still open). Decide layering when taken up (peer workflow vs. its own repo) — likely via /brainstorming. Surfaced 2026-06-08 puppeting the FFXI RE target. (Backburnered 2026-06-09 — puppet workspace untouched; this is just the parked bookmark.)
  • 2026-06-10 — Adopt find-skills' quality-verification heuristics in the learning-new-skills skill: weight surveyed patterns by signals of being battle-tested vs. fringe — install count (e.g. ≥1K), source reputation (vercel-labs / anthropics / microsoft treated as more trustworthy), and GitHub stars. Use them as a filter when mining skills.sh + other registries for "ideas worth stealing," not as install gating (the skill never installs anything). Source: Vercel's find-skills skill (skills.sh ecosystem), examined 2026-06-10. (Backburnered 2026-06-10 — Brad liked it but parked it while just adding skills.sh as the primary source for now.)
  • 2026-06-10 — Two new RE disciplines to expand Odin Codin' beyond DAT-format RE (both surfaced 2026-06-10; the full RE stack is out of scope for now). (1) Binary/DLL machine-code RE, Ghidra-MCP driven — reverse-engineer the FFXI client .exe/.dll via Ghidra (free decompiler) or radare2 through an MCP bridge (e.g. LaurieWired/GhidraMCP, bethington/ghidra-mcp — exposes ~200 RE tools to Claude; requires the tool installed). This is a different discipline from DAT-format RE (machine code, not data) — and NOT vgrichina/re-skill's hand-rolled disassembler, which only suits simple retro CPUs. Head start: the FFXI client-RE community — aikar/SigScan (runtime signature scanning), Ashita/Windower (already map much of the client), atom0s/XiEvents (event/cutscene VM). Payoff: the client code is the ground-truth oracle for DAT formats — read the actual parser the client ships instead of triangulating POLUtils/xi-tinkerer/ShiningFantasia. Caveats: legal/ToS (live-MMO client RE is modding-community territory, not SE-sanctioned — ties to the LSB social-license item); a large, decades-old binary. (2) Packet-crawl mode — ingest a Kouseki-style capture (PacketViewer + ActionView/CapLog) against atom0s/XiPackets (the per-opcode struct "Rosetta Stone") to emit mob/behavior SKILL.md specs for private-server implementation. Proof-of-concept done this session: working/research/specs/hnm-fafnir-immunities/SKILL.md. Both are "crawl → AI-facing SKILL.md" Odin Codin' modes — siblings to the existing "build a proper FFXI client-RE project" backburner item. Surfaced 2026-06-10 in a packet-RE feasibility exploration.
  • 2026-06-02 — Learning project (fun, on-theme). Build a generic personal learning workflow around Yggdrasil: the plan for how Brad researches and learns new skills, tracks his progress, and keeps up with current developments — for any subject, not just AI. Reusable pieces likely include quizzing/testing to surface gaps, progress tracking, and a "keep current" research loop (note the overlap with the existing learning-new-skills skill, which keeps the toolkit current — different target, related machinery; reconcile when designing). "Learning AI" is the first implementation/instance of this generic project: quiz Brad on what he's actually been building with (skills, hooks, context windows, memory, allowed-tools/disable-model-invocation, MCP, slash commands), surface gaps, and produce a structured AI curriculum. Strong fit because Yggdrasil is the domain-agnostic plan/design partner — a learning curriculum is a non-code plan instance, and this is an ideal first real consumer of the planning skill (low-stakes, motivating dogfood). Open layering question for the brainstorm: generic learning project as a peer workflow at the Odin Codin' layer (reusable learning skills) with AI-learning as a plan instance in its own repo — or some blend? Could pull the evolving "what Brad knows" inventory out of CLAUDE.md (deliberately kept lean there 2026-06-02) into the AI-learning instance's own living doc. Update 2026-06-10: the stated precondition — "bring to /brainstorming after the planning skill exists" — is now met (planning skill shipped 2026-06-08); ready for a /brainstorming pass whenever it warms up. (Backburnered 2026-06-10 — a loose idea, not of active interest right now; per Brad. Adjacent in spirit to the journaling-agent item above.)