Mother skill for platform-descriptor sister skills. Defines what a BIOME is โ a bounded region of an ecosystem (coexisting, exchanging, never isolated) for one platform you operate โ and what files, subdirectories, and cross-biome bridges every daughter biome inherits.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Mother skill for platform-descriptor sister skills. Defines what a BIOME is โ a bounded region of an ecosystem (coexisting, exchanging, never isolated) for one platform you operate โ and what files, subdirectories, and cross-biome bridges every daughter biome inherits.
MOOLLM โ lifted from three sister biomes that accreted independently in a consumer repo (leela-ai/central: edgebox, gcs, azure) then congealed into a publishable meta-skill pattern, 2026-04
BIOME โ Mother Skill for Platform Descriptors
BIOMEn. โ a bounded region of an ecosystem, with its own inhabitants and climate, coexisting with sister biomes and exchanging matter, energy, and information with them along defined gateways.
This skill is the mother prototype that platform-descriptor sister skills inherit from. Each daughter biome declares inherits: biome in its CARD.yml / GLANCE.yml / SKILL.md frontmatter, and its directory structure mirrors the skeleton this skill defines.
Six example biomes ship with this skill for reference, each illustrating a distinct dimension:
Multi-role โ CDN + Workers + R2 + DNS + Tunnels + Pages; the "many simultaneous roles" lesson
Consumer repos adopt this skill and populate their own biomes/registry.yml + biomes/gateways.yml + per-biome daughter skills in their own repos. The example biomes here are vendor-neutral reference material, not a working registry.
Why "biome" โ not "microworld," "virtual world," or "platform"
The name is deliberate. The ecological framing does three things no other candidate does:
Microworld (Papert's term, MOOLLM's ambient usage) implies an space โ a pedagogical sandbox sealed against distraction. A MOOLLM microworld is single-subject by design. Biomes are explicitly โ inhabitants cross boundaries; a cloud MIG pulls images from the github biome; an edge box pulls models from a cloud biome; tailscale threads SSH through all of them at once. Isolation is the wrong shape.
isolated
not isolated
Macroworld / virtual world implies simulation or containment. Biomes are real operational territory โ they aren't modeled, they're inhabited. Machines run there, bills come from there, outages happen there.
Platform is accurate but flat. "Platform" hides that there's weather, seasons, predators, migrations โ that things live in these places, have needs, interact. The platform-language invites thinking of GCP and Azure as interchangeable slots. The biome-language makes it obvious that the Amazon is not the Sahara is not a coral reef, even if they all "have climate and elevation."
The ecology metaphor earns its keep when you start looking at cross-biome flows:
Exchange โ containers built in github, consumed by cloud VMs; videos produced at edge boxes, flowing up to cloud buckets; models trained somewhere, distributed everywhere.
Symbiosis โ tailscale overlays every biome without owning any of them; a devops meta-biome serves images to every biome and exists for no other purpose.
Climate โ each biome has its own rate constants: cloud MIG autoscaler cooldowns, edge spot preemption rates, CI runner queue times.
Niches โ each biome supports workloads the others can't (or won't): edge runs local because it has to, cloud runs pay-per-use because it can, forges run CI because they already have the code.
All of these are captured structurally by the skeleton (recipes/instances/runbooks/gotchas/...) and by the consumer repo's biomes/gateways.yml, which is the explicit record of cross-biome exchange. Gateways are not a metadata afterthought โ they're the biological substance of the ecosystem.
Structural ancestry. This pattern is directly modeled on MOOLLM's schema skill (schemapedia). Schemapedia maintains a registry of schema mechanisms (json-schema, sqlite, git, drescher, โฆ) with registry.yml + plugin-convention.yml + gateways.yml. The biome pattern applies the same idea one level up โ each biome is a full sister skill rather than a plugin entry inside a single skill, but the registry / convention / gateways triad is deliberately parallel. See biomes/convention.yml and biomes/families.yml.
Inheritance, the Self way
Each daughter biome is an instance of the biome prototype:
# In skills/gcs/CARD.yml (and GLANCE.yml)inherits:-biome
When a daughter biome doesn't define a convention locally, the answer lives here โ first-match-wins delegation per MOOLLM's Delegation Object Protocol (DOP). When you land in a daughter biome and the question is "what does this skeleton mean?", consult this skill.
Why a shared skeleton
Predictable navigation. Once you know where to find gotchas in one biome, you know where to find them in any biome.
Low-friction expansion. New biomes (runner, tailscale, future platforms) grow into the skeleton instead of inventing their own.
Cross-biome learning. Lessons from one biome (e.g., "GCP has invisible quota for new accelerator families") find their parallels in another (e.g., "Azure has 0-default quota for new VM families") because both live at the same path.
MOOLLM alignment. Each biome is a microworld โ a bounded, coherent space where platform-specific vocabulary and patterns make sense. The skeleton is the shared grammar.
What makes a biome
A MOOLLM-compatible platform biome is a directory under skills/ that describes one operational platform โ a place where you run workloads (cloud VMs, on-prem hardware, CI runners), a registry or image source (Artifact Registry, Docker Hub, a private package index), an identity or network plane you depend on (Tailscale, Cloudflare, Firebase), or any other bounded infrastructure surface whose specific vocabulary and quirks are worth capturing.
Not every skill is a biome. A biome is specifically a platform descriptor โ it encodes operational knowledge of a bounded infrastructure surface. Non-biome skills include process skills, workload skills, and meta-skills (e.g., skill, card, cauldron, schema, file-system-object). A biome declares membership by inheriting from this skill.
Core files (required โ the Semantic Image Pyramid)
Protocol overview, core mental model, worked examples, documentation map
300-500 lines
README.md
Human landing page with reading order and quick links
50-80 lines
Core subdirectories (standard skeleton)
Every biome should have these directories, even if initially populated only by a README.md stub. They fill in over time as we learn.
Directory
Contents
Reading convention
recipes/
Ready-to-run commands and short patterns. Markdown files organized by task theme. The "bash snippet you Google for at 2am" section.
recipes/<topic>.md; plus recipes/README.md as index
instances/
Per-deployment knowledge cache. One subdir per deployed instance (project, subscription, on-prem box). Contains state.yml, history.md, per-instance deploy.md where useful.
instances/<name>/README.md
runbooks/
Task-oriented walkthroughs. Each runbook covers one multi-step operation end-to-end.
runbooks/<UPPERCASE-TASK>.md
gotchas/
Known failure modes, one per file. Each includes the symptom, the root cause, and the fix. The archive of "this took a deploy cycle to figure out."
gotchas/<kebab-case-symptom>.md
protocols/
Deep documentation of mechanisms and invariants. Longer-form than recipes, more stable than gotchas.
protocols/<topic>.md
scripts/
Automation hooks specific to this biome (probes, diagnostic wrappers). Not user-facing CLIs (those live in tools/).
scripts/<name>.sh
cost/
Date-stamped billing analyses. Each file captures a specific finding with the queries, raw numbers, interpretation, and action.
cost/YYYY-MM-DD-topic.md
Cloud-biome-specific subdirectories
For biomes whose platform is a public cloud (gcs, azure, future aws):
Directory
Contents
quota/
Quota inventory per region/family. What we have, what we've requested, what's blocking.
regions/
Region/zone/SKU availability catalogs. "Where does our fleet run, and why."
Edge-biome-specific subdirectories
For edgebox (and any future terrestrial biomes like customer-hardware):
Directory
Contents
stacks/
Stack types (vision, builder, training). A biome-specific concept โ an edgebox can be deployed as different stack shapes; a GCS project can't.
templates/
Template files for new instances.
CLI tool ecosystem (optional but encouraged)
Every biome operates through a set of command-line tools โ gcloud for GCP, az for Azure, aws for AWS, ssh/docker/systemctl for edge/on-prem, gh/git for forges, tailscale for mesh network, wrangler/flarectl for Cloudflare. The biome's CARD.yml should list them in a cli_tools: block so newcomers (and LLMs) can tell at a glance "which binary do I invoke for which concern."
cli_tools:-name:gcloudrole:"GCP CLI for compute, secrets, IAM, pubsub, etc."install:"brew install --cask google-cloud-sdk"auth:"gcloud auth login --no-launch-browser"primary_usage:"recipes/gcloud-patterns.md"vendor_docs:https://cloud.google.com/sdk/gcloud-name:gsutilrole:"GCS object storage CLI (bucket listings, cp/mv, lifecycle)"primary_usage:"bucket reads/writes, syncing artifacts"also_used_by: [edge-on-prem] # shared CLI โ record as a gateway
Tools used across multiple biomes (docker, git, ssh, gsutil, tailscale) should be listed in every biome that uses them AND get a corresponding bridge entry in the consumer repo's biomes/gateways.yml with kind: shared_cli_tool.
Optional / grows-as-needed
Directory
When
examples/
Free-form examples not fitting into recipes/runbooks.
Use templates/BIOME/ as the directory scaffold to instantiate from when adding a new biome.
When to promote content into the skeleton
Content lives in one of three places:
SKILL.md โ the canonical overview, everything cross-referenced.
Subdirectory file โ when a topic grows beyond 1-2 paragraphs in SKILL.md, or when it has its own lifecycle (date-stamped cost analyses, per-instance state, gotchas with distinct identity).
External doc (under a repo-level docs/ path or a sibling skill) โ when the topic spans multiple biomes or is cross-cutting (e.g., a cloud-gpu-platforms.md that catalogs GCP + Azure + AWS GPU SKUs side-by-side).
The skeleton establishes where things can live. Movement is lazy: keep it in SKILL.md until it earns its own file by being longer than a paragraph or having an independent lifecycle.
CDN + Workers + R2 + DNS + Tunnels + Pages; one brand, many roles
Each of these is a stub with the minimal skeleton (GLANCE.yml, CARD.yml, README.md, and a seed README.md per required subdirectory). Consumer repos are expected to either:
(a) fork the stub into their own skills/<biome>/ and grow it with site-specific recipes/instances/gotchas, declaring inherits: biome; or
(b) leave the stub as-is in moollm/ and write a thin consumer-side overlay skill that inherits from it (same inherits: biome chain, plus inherits: biome.<biome-name> for delegation into the stub's CLI tool list / vendor doc links).
Candidate future example biomes
Platforms common enough to deserve a reference stub when someone authors one:
Candidate
Nature
Why it matters
firebase
Auth / realtime DB / hosting
Multi-role like Cloudflare; auth plane crosses every biome that consumes identity
postgres
Database
Spans managed (RDS/Cloud SQL/Azure DB) + self-hosted; worth documenting both
bigquery
Data analytics
Cloud data-warehouse shape; sister to Snowflake/Redshift
kubernetes
Orchestrator-as-platform
When consumer fleets include k8s; distinct enough to deserve its own biome
terraform
IaC
Cross-cutting like github โ every biome consumes it
1password / vault
Secrets plane
Identity-adjacent overlay biomes
New example biomes become worth shipping when they illustrate a distinct biome shape (new family, new cross-cutting pattern, new failure-mode cluster) โ not just another cloud vendor.
Parallel to schemapedia โ which files do what
Schemapedia and biome-pedia both split responsibility across a registry / convention / gateways triad. Side-by-side:
Each biome's subdirectories (recipes/, runbooks/, gotchas/, โฆ) are native MOOLLM augmentations, indexed via the biome's own CARD.yml
Key structural difference: schemapedia members are small plugin entries that fit inside one skill directory; biome members are full skills with their own sub-skeletons and Self-style inheritance. The registry is correspondingly thinner (it points at sibling skill directories), and each biome carries significantly more internal structure than a MECHANISM.yml does. But the role of each file is the same.
Relationship to MOOLLM
Each biome is a microworld in Seymour Papert / MOOLLM sense โ a bounded space with its own coherent vocabulary and rules, where learning the rules of one biome doesn't require learning the rules of any other. The shared skeleton is the grammar that spans microworlds: you can learn the grammar once and read any biome fluently. The skeleton gives you sisterhood across microworlds without collapsing them into one undifferentiated vendor-neutral mush.
MOOLLM concepts this skill embodies:
Skills as prototypes (Self / DOP) โ biome is a prototype; gcp, azure, aws, github, tailscale, cloudflare are instances declaring inherits: biome. When a daughter doesn't define a convention, lookup falls back to this skill.
Prototypes and instances โ the daughter biome is itself a prototype for ITS instances (e.g., gcp/instances/<project-id>/).
File System as Object โ plural directory names declare element types; UPPERCASE marker files declare interface exports. See file-system-object.
K-lines as filenames โ directory and file names are activation handles.
PLAY-LEARN-LIFT โ content starts in chat/SKILL.md (play), gets documented (learn), gets promoted into the skeleton with cross-references (lift). This skill itself is a LIFT from three concrete biomes that accreted independently in a consumer repo before the pattern became visible.
Sister scripts โ code stays generic across biomes, values vary (e.g., <TYPE>_VM_* secret keys are sister keys; same shape, cloud-appropriate contents).
How to add a new biome
Decide it's a biome. A biome is a platform descriptor โ a place you operate, not a process or workload skill. If unsure, list it under candidates: in your consumer repo's biomes/registry.yml instead of starting a directory.
Pick a name + family. Family is one of cloud, edge, meta, overlay, collaboration, execution. See biomes/families.yml.
Fill in the placeholders (name, description, family, character). You may also start from one of the example biomes shipped here (copy biomes/gcp/ as the cloud shape, biomes/github/ as the cross-cutting shape, etc.) and edit.
Declare inheritance. Each of <new-biome>/CARD.yml, <new-biome>/GLANCE.yml, and <new-biome>/SKILL.md (frontmatter) gets inherits: [biome]. This makes the prototype chain explicit and machine-detectable. If inheriting from an example biome here, also add inherits: [biome, biome.<example-name>] so delegation walks into the stub first before falling back to the mother.
Register (in your consumer repo). Add an entry to your biomes/registry.yml under the right family.
Cross-biome bridges. If the new biome shares machinery with existing biomes (secret naming, shared scripts, image distribution, mesh networking, DNS ownership, โฆ), add bridge entries to your biomes/gateways.yml.
Grow. Stub README.md in each subdirectory; populate as operational knowledge accumulates. Promote content out of <new-biome>/SKILL.md into the appropriate subdirectory when it grows beyond 1-2 paragraphs or has its own lifecycle.
How to grow a biome (general operating loop)
New operational finding? Put it in <biome>/SKILL.md first.
Finding grows beyond a paragraph or is one of a series? Promote to the right subdirectory with a dedicated file. Add a line to the subdirectory's README.md index.
Finding applies across biomes? Extract to a repo-level doc or a sibling skill; have each biome's SKILL.md link to it. Add a bridge entry in your biomes/gateways.yml.
Subdirectory is empty because you haven't hit that kind of thing yet? Leave the README.md stub. It's a signpost to your future self about what you'd eventually want to capture there.
Cross-cutting biomes (the github / tailscale / cloudflare pattern)
Some biomes are cross-cutting by design โ they don't just coexist with other biomes, they pass through every other biome:
github โ every cloud biome pulls deploy artifacts from GitHub Actions; every edge biome is reached via git pull; every infrastructure change is audited there. The forge IS the circulatory system.
tailscale โ every SSH session, every private-service probe, every cross-region debug traverses the mesh. Tailscale overlays all biomes without owning any of them.
cloudflare โ DNS for every biome's public-facing name, CDN in front of web biomes, Workers/R2 as compute+storage, Tunnels reaching back into private biomes. Multi-role by design.
Secrets plane (e.g., 1password, vault) โ same pattern; would deserve its own cross-cutting biome stub when captured.
Cross-cutting biomes are structurally the same shape as bounded ones (same skeleton, same required files) but semantically the substance of the gateways themselves. When you write biomes/gateways.yml in a consumer repo, many bridges will be the cross-cutting biome โ e.g., "edge โ cloud via tailscale" is a gateway entry whose kind: mesh_overlay points at the tailscale biome.
file-system-object โ the grammar daughters use (UPPERCASE marker files = interface exports)
yaml-jazz โ semantic YAML used in convention/registry/gateways files
sister-script โ the cross-biome scripting pattern that lets one script serve every biome
Publication
This skill is Tier 1 (MOOLLM core). It is published as part of the MOOLLM distribution and is freely reusable under MIT. Consumer repos adopt it by declaring inherits: biome in their platform-descriptor skills and populating their own biomes/registry.yml + biomes/gateways.yml.
Origin: this skill was synthesized from three sister biomes (edgebox, gcs, azure) that accreted independently in a consumer infrastructure repo (leela-ai/central), until the shared shape became obvious enough to extract. The PLAY-LEARN-LIFT arc โ three concrete biomes โ meta-skill โ is the archetypal MOOLLM lift.
