Understanding Agentic Guidelines

Human-facing explanation of how the Agentic Guidelines page works — its structure, language rules, and how to keep it accurate. A how-to reference, not an open invitation to edit.

What the Agentic Guidelines Page Is

The parent page — Agentic Guidelines — is not a documentation page in the normal sense. It contains no prose meant for a human to read comfortably. It is a structured seed document, intended to be pasted or loaded into an AI agent session before any GS_Play framework work begins.

Think of it as a mission briefing card. It is dense, compressed, and declarative. Every line exists to be parsed and acted on by a machine — not to be understood leisurely by a person.

Why it exists: Agents working without this document make predictable, repeatable errors. They guess EBus names incorrectly, place EnableForAssetEditor in the wrong reflection context, write polymorphic containers with the wrong pointer type, or generate C++ where a scripting answer was needed. The agentic guidelines page is the countermeasure for all of those failures.

What it is not: It is not a tutorial, an overview, or a feature summary. Those documents live elsewhere in the framework documentation. This page exists only to give an agent precise, actionable constraints and lookup data before it starts working.


Who Uses This Framework and Why It Matters

The framework has two distinct user groups. This distinction determines what kind of guidance an agent should give, and it is the single most important thing to keep in mind when editing this document.

Designer / Scripter

This user works entirely inside the O3DE editor. They configure components in the entity inspector, connect logic with Script Canvas or Lua, set up prefabs, and tweak exposed properties. They never write C++. For this user, an agent should provide guidance about:

  • Which components to add to an entity
  • How to configure component properties
  • How to wire EBus events through Script Canvas nodes
  • How to structure prefabs and entity hierarchies

Engineer / Developer

This user writes C++ to extend the framework. They inherit from GS_Play base classes, create new gems or project-level systems, write new components, and build on top of the existing architecture. For this user, an agent should provide guidance about:

  • Which base class to inherit from
  • How to implement EBus interfaces
  • How to register new types (dialogue effects, pulse types, etc.)
  • How to structure reflection, module registration, and CMake targets

Engineers have full freedom to combine gems. GS_Complete is the framework’s reference integration layer — it demonstrates cross-gem patterns but is not a restriction on what engineers can do in their own projects. An engineer may freely include headers from multiple gems in their project code.


How the Document Is Structured

The page uses three tiers of information. Understanding these tiers is the most important thing before contributing.

Tier 1 — Invariants

Hard rules with no exceptions. These live at the very top of the document in the INVARIANTS section. An agent that reads only the first 600 tokens of the document will encounter all of them.

Invariants are:

  • Specific, not general (“polymorphic containers use BaseT*”, not “use the right container type”)
  • Complete enough to act on without reading anything else
  • Framed as prohibitions or exact requirements — never suggestions

Do not add to the Invariants section unless a rule is a hard technical constraint that has no contextual exceptions and causes a silent failure or editor malfunction when violated.

Tier 2 — Ontology and Reference

This is the structural description of what the system is: the SYSTEM ONTOLOGY paragraph, the GEM INDEX table, the EBUS DISPATCH REFERENCE, and the HOT PATHS section. This tier answers the question: what is this system and how do I operate it?

This tier is:

  • Table-first. Prose is minimal.
  • Accurate. A single wrong EBus method name here means an agent fails silently.
  • Compressed. One row per concept, not one paragraph per concept.

Tier 3 — Conditional Depth

This tier does not provide information directly. It tells the agent when to stop and go deeper. The CONTEXT ANCHORS, CONFIDENCE THRESHOLDS, and CLARIFICATION TRIGGERS sections all belong here. They point the agent toward more information and define the exact conditions under which to seek it.


The Token Grammar

The document uses a small, fixed vocabulary of bracketed tokens. These act as machine-readable instruction labels. Agents across all backends pick them up reliably through pattern matching.

TokenMeaningLocation
[INVARIANT]Hard constraint. Never violate. No exceptions.INVARIANTS section
[ANCHOR]Stop and read the linked documentation before proceeding.CONTEXT ANCHORS, inline in HOT PATHS
[MEMORY]Persist this content if a memory system is available.MEMORY PROTOCOL section
[ASK]Do not infer. Pause and ask the user before proceeding.CLARIFICATION TRIGGERS, inline in HOT PATHS
[ANTIPATTERN]What not to do and briefly why.ANTIPATTERN CATALOG section

Use only these tokens. Do not invent new ones. Tokens work because they are stable — an agent learns to treat [INVARIANT] as a certain category of instruction. A new, undiscussed token may be ignored or misinterpreted.

If you genuinely believe a new token category is needed, document the reasoning and discuss it before adding it. The bar is high: the existing five tokens cover almost every instruction type needed.


Editing the Agentic Guidelines: Step-by-Step Workflow

Step 1 — Identify what changed

Before opening the file, answer these questions:

  • Was there a codebase change (new EBus, new component, renamed method)?
  • Was there an error an agent actually made — something it got wrong that this document should have prevented?
  • Is there a new common task that agents are regularly asked to do with no hot path covering it?
  • Is there a new documentation page that an agent should be directed to?

If none of these are true, you probably do not need to edit the parent document. Editing it without a concrete reason adds noise.

Step 2 — Locate the right section

Use this map to find where your change belongs:

Type of changeTarget section
A rule that must never be brokenINVARIANTS
A new gem or renamed managerGEM INDEX
A new or renamed EBus or methodEBUS DISPATCH REFERENCE
A new common code taskHOT PATHS
A mistake agents keep makingANTIPATTERN CATALOG
A decision that needs user inputCLARIFICATION TRIGGERS
A documentation page that gives deep contextCONTEXT ANCHORS
Something agents can safely assumeCONFIDENCE THRESHOLDS
A memory entry agents should persistMEMORY PROTOCOL

If your change doesn’t fit any of these cleanly, it is probably better expressed in the framework’s technical documentation pages rather than here.

Step 3 — Apply the minimum change

Do not add surrounding context, explanations, or improvements to adjacent content. The agentic document is optimized for token density — adding prose that an agent does not need to act on degrades its quality.

Apply the minimum correct change:

  • One new row in a table
  • One new [INVARIANT] entry
  • One revised method name
  • One new numbered step in a hot path

Resist the impulse to expand.

Step 4 — Verify accuracy

Before saving, verify every name, method, and path in your change against the source:

  • EBus method names — Open the source header for that bus. Copy the method name exactly. Do not paraphrase.
  • Doc paths — Confirm the path actually exists. A broken anchor is worse than no anchor.
  • Base class names — Confirm against the gem’s public Include/ headers.
  • TypeIds — Do not add TypeId values unless you have copied them from GS_{Name}TypeIds.h.

Step 5 — Apply the self-test

Read only the section you changed and ask: does an agent reading only this section know what to do, or does it still need to infer?

If it still needs to infer, the entry is incomplete. If it specifies exact names, exact locations, exact conditions, and exact patterns — it is ready.

Step 6 — Update this contributing page if structure changed

If you added a new section to the parent document, added a new token type, changed the tier structure, or changed any guidance about how to contribute — update this page to match. The two documents must stay in sync. See the section below: When to Update This Page.


Section-by-Section Editing Reference

BOOTSTRAP PROTOCOL

Edit when: The loading order or session re-entry behavior of the document should change.

Rules:

  • Keep this section under 8 lines. It is the first thing an agent reads.
  • Instructions must be ordered. Numbered list only.
  • Do not add explanatory prose. Each line is an action, not a description.

INVARIANTS

Edit when: A new silent-failure rule has been discovered through a real agent error, or an existing invariant is no longer accurate.

Qualifying test for a new invariant — all must be true:

  1. It is a specific technical rule, not a general principle.
  2. Violating it causes a failure that is silent or extremely hard to diagnose.
  3. It applies in all contexts, with no exceptions.
  4. It cannot be reliably derived from general O3DE knowledge.

Format:

`[INVARIANT]` **Short title.** One or two sentences of exact, actionable constraint.
Optionally: a code block showing the correct pattern.

Do not add O3DE general best practices here. They do not belong in the invariants list — agents already know them.


MEMORY PROTOCOL

Edit when: A new section is added to the parent document that should be persisted to memory, or a memory key is renamed.

Rules:

  • Memory keys are lowercase with underscores: gs_play_invariants, not GS Play Invariants.
  • Only add memory entries for sections that are dense enough to be worth re-loading from memory rather than re-reading.
  • The freshness rule (source code wins over stored memory) must never be removed.

SYSTEM ONTOLOGY

Edit when: The engine version changes, the build toolchain changes, or the platform target changes.

Rules:

  • The attribute table is for hard facts only: version numbers, language, build system, platform.
  • The “Core architectural rules” bullets must remain at three or fewer. If you need to add a fourth, one of the existing three should be absorbed into a more general statement.
  • Do not add GS_Complete or internal gem restrictions as architectural rules. These are framework-internal development constraints, not user-facing rules.

GEM INDEX

Edit when: A new gem is added, a gem is renamed, a manager component is added or renamed, a primary bus is renamed, or a documentation path changes.

Rules:

  • One row per gem. Do not split a gem across two rows.
  • The dependency graph must be updated any time the gem dependency structure changes.
  • Doc paths must be real paths. Verify before adding.
  • GS_Complete remains in the index as the integration layer reference — it documents a real gem. Its description should clarify it is Genome Studios’ reference layer, not a restriction on user code.

EBUS DISPATCH REFERENCE

Edit when: A new EBus is added to any gem, a method is renamed, or the dispatch policy of a bus changes.

Rules:

  • Method names are copied from source headers. Not paraphrased, not summarized — exact.
  • Include only the most commonly-called methods. This is a quick-lookup reference, not a full API.
  • If a bus is entity-addressed, the Dispatch column must say ById. If it is a singleton broadcast, the column says Broadcast. If it is a listener-only bus, say Listen.
  • Adding a new gem requires a new ### {GemName} buses subsection.

HOT PATHS

Edit when: A new common task emerges that agents are regularly asked to perform, or an existing hot path’s steps become inaccurate.

Qualifying test for a new hot path — all must be true:

  1. Agents are asked to do this task regularly.
  2. There is a correct pattern that is non-obvious or differs from standard O3DE conventions.
  3. There is a genuine way to get it wrong that the hot path would prevent.

Format:

### HOT PATH N — Short task description

1. First discrete action.
2. Second discrete action.
3. Third discrete action.

`[ANCHOR]` Condition requiring deeper reading → /docs/path/to/page/

Rules:

  • Steps are single, discrete actions. Not compound sentences.
  • No prose between steps.
  • Add an [ANCHOR] at the end if there is a documentation page that provides necessary deeper detail.
  • Hot paths are for engineer (C++ extension) tasks. Designer tasks belong in component or scripting documentation, not here.

ANTIPATTERN CATALOG

Edit when: A real agent error is identified that this document should prevent, or an existing antipattern is no longer relevant.

Qualifying test — all must be true:

  1. An agent actually made this mistake, or will predictably make it without the warning.
  2. The failure mode is silent or ambiguous — not a compile error that would surface immediately.
  3. The correct alternative is already documented (in INVARIANTS or HOT PATHS).

Format:

`[ANTIPATTERN]` **Short title.** One sentence describing the failure. One sentence pointing to the correct behavior or its location in this document.

Do not add antipatterns that are just general bad practices. Every entry here should represent a failure mode specific to GS_Play or to O3DE patterns that agents commonly get wrong.


CONFIDENCE THRESHOLDS

Edit when: A GS_Play-specific assumption has been confirmed safe (add to “proceed” list) or a pattern that agents incorrectly assume they know turns out to need verification (add to “stop” list).

Rules:

  • The “proceed” list contains patterns where O3DE general knowledge is accurate and sufficient.
  • The “stop” list contains patterns where GS_Play diverges from general O3DE knowledge, or where API details are too precise to guess.
  • Do not add items to the “proceed” list unless they have been verified accurate across the framework.

CONTEXT ANCHORS

Edit when: A new documentation page is added that meaningfully fills a knowledge gap an agent would otherwise have to guess around, or an anchor’s target path changes.

Qualifying test for a new anchor:

  1. There is a specific documentation page at a specific path.
  2. An agent working without that page would make structural errors on this task.
  3. The condition is narrow enough to fire at the right moment (“before writing unit movement code”) — not so broad that it triggers on general O3DE work.

Format:

`[ANCHOR]` **Before doing [specific task]:**
→ Read /docs/path/to/page/ for [what the agent needs from it — be specific].
→ `[ASK]` Optional inline clarification prompt if a user decision is required here.

Note on GS_Complete anchors: GS_Complete is a valid reference target. An anchor pointing to it as a reference (“review GS_Complete for examples”) is appropriate. An anchor that frames it as a restriction (“you must implement this there”) is not — that is a framework-internal structural convention, not a rule that applies to user project code.


CLARIFICATION TRIGGERS

Edit when: A new category of user decision is identified where agents predictably infer the wrong answer, or an existing trigger is too broad to fire usefully.

Qualifying test:

  1. The agent cannot determine the correct answer from documentation alone — it requires user intent.
  2. Agents predictably get this wrong when left to infer.
  3. Getting it wrong has real consequences (wrong code generated, wrong scope, irreversible decisions).

The USER CONTEXT trigger at the top of this section must never be removed. It is the most important clarification in the document — it gates all subsequent behavior between designer/scripter guidance and engineer guidance.

The two documentation sections referenced in USER CONTEXT are:

  • The Basics (/docs/the_basics/) — Scripter and designer documentation. Covers component usage, property configuration, and scripting for all gems.
  • Framework API (/docs/framework/) — Engineer documentation. Covers C++ interfaces, EBus signatures, base class architecture, and reflection patterns for all gems.

Both sections cover the same featureset across the same gem set: core, cinematics, environment, interaction, performer, phantomcam, ui, unit, ai, item, rpstats. Update the USER CONTEXT doc links if any gem is added, renamed, or its documentation path changes in either section.


What Makes Bad Agentic Content

Understanding what to avoid is as important as understanding what to write.

Prose explanations. The parent document must never read like a tutorial. “The EBus system in O3DE works by…” is wasted tokens. An agent already knows what EBus is. It needs the specific name and specific dispatch syntax.

Vague headings. “Common Patterns”, “Unique Details” — these tell neither a human nor a machine what type of instruction to expect. Every section heading in the parent document is also a category of instruction.

Soft language. “You should consider”, “it is generally recommended” — agents do not weight soft language correctly. If something must be done a specific way, say so directly. Preferences belong in other documentation pages, not here.

Information without a trigger condition. Raw information with no instruction about when to use it adds cognitive load without adding value. Every fact in the parent document is attached to a pattern, a constraint, a lookup trigger, or a warning.

Framework-internal conventions framed as user-facing rules. Conventions about how the framework’s own gem structure is organized (like where cross-gem components live within the framework itself) should not appear as [INVARIANT] or [ANTIPATTERN] entries. Engineers extending the framework in their own projects have full freedom to organize code as their project requires.

Outdated EBus names or method signatures. This is the most damaging type of error in the document. An agent trusting a wrong bus name fails silently, and the diagnosis will be non-obvious. Before adding or updating anything in the EBus section, verify against the current source header.


Maintaining Accuracy

When any of the following changes in the codebase, update the corresponding section:

Codebase changeSection to update
New EBus addedEBUS DISPATCH REFERENCE — add table row
EBus method renamedEBUS DISPATCH REFERENCE — update row; scan HOT PATHS for any direct references
New gem addedGEM INDEX — new row; update dependency graph
New Manager Component added or renamedGEM INDEX — update Manager column and Primary Bus column
New documentation page publishedCONTEXT ANCHORS — add anchor if it fills a knowledge gap
Documentation page moved or renamedCONTEXT ANCHORS — update the path; if it is a gem root page in The Basics or Framework, also check CLARIFICATION TRIGGERS USER CONTEXT links
New gem added to the frameworkGEM INDEX, EBUS DISPATCH REFERENCE, CLARIFICATION TRIGGERS USER CONTEXT gem list (both /docs/the_basics/ and /docs/framework/ paths)
O3DE API change affecting Reflect() patternsINVARIANTS and HOT PATH 1, 5 — verify code blocks still compile
New common agent task identifiedHOT PATHS — add new numbered path
New silent-failure pattern discoveredANTIPATTERN CATALOG — add entry
GS_Play base class API changesCONFIDENCE THRESHOLDS — verify the “stop” list is still accurate

When to Update This Contributing Page

This contributing page should stay in sync with the parent document’s structure. Update it when:

  • A new section is added to the parent document — add a corresponding entry to the Section-by-Section Editing Reference above.
  • A token is added to the grammar — add it to the Token Grammar table.
  • The tier structure changes — update the “How the Document Is Structured” section.
  • The distinction between user contexts changes (new user type, renamed category) — update “Who Uses This Framework”.
  • Any “rules” in the section-specific guidance become outdated.

This page is allowed to have prose. It is written for humans. Explanations, reasoning, and examples are all appropriate here. The parent document is the one that must remain dense and machine-optimized.


Testing Your Contribution

Before saving any change to the parent document, apply this test:

Read only the section you changed. Ask: does an agent reading only this section know exactly what to do, or does it still need to infer?

If it still needs to infer, the entry is incomplete. It is ready when it specifies:

  • Exact names (method names, class names, bus names)
  • Exact locations (file paths, doc paths, section names)
  • Exact conditions (the trigger that fires this instruction)
  • Exact patterns (the correct code form or step sequence)

A strong contribution eliminates a whole class of agent errors. A weak contribution adds words without eliminating uncertainty. When in doubt, add nothing and write the knowledge into the relevant framework documentation page instead. Let a [ANCHOR] point to it.