diff --git a/src/bub/agent/core.py b/src/bub/agent/core.py
index fc5bb55f..b1c648ae 100644
--- a/src/bub/agent/core.py
+++ b/src/bub/agent/core.py
@@ -10,6 +10,7 @@
 from republic import LLM, Tool
 
 from ..errors import ModelNotConfiguredError
+from ..skills import build_skills_prompt_section
 from .context import Context
 
 MODEL_NOT_CONFIGURED_ERROR = "Model not configured. Set BUB_MODEL (e.g., 'openai:gpt-4o-mini')."
@@ -105,11 +106,20 @@ def respond(
         on_event: Callable[[ToolEvent], None] | None = None,
     ) -> str:
         """Run a tool-aware response and return assistant text."""
-        if self._system_prompt:
-            messages = [{"role": "system", "content": self._system_prompt}, *messages]
-
+        inserted_system = False
         observation_fingerprints: dict[str, str] = {}
         while True:
+            system_prompt = self._build_system_prompt()
+            if system_prompt:
+                if inserted_system:
+                    messages[0] = {"role": "system", "content": system_prompt}
+                else:
+                    messages = [{"role": "system", "content": system_prompt}, *messages]
+                    inserted_system = True
+            elif inserted_system:
+                messages = messages[1:]
+                inserted_system = False
+
             response = self._llm.chat.raw(
                 messages=messages,
                 tools=self._tools,
@@ -134,6 +144,13 @@ def respond(
 
             return text or ""
 
+    def _build_system_prompt(self) -> str:
+        base_prompt = self._system_prompt.strip()
+        skills_section = build_skills_prompt_section(self._context.workspace_path)
+        if not skills_section:
+            return base_prompt
+        return f"{base_prompt}\n\n{skills_section}"
+
     def _execute_tools(
         self,
         tool_calls: list[dict[str, Any]],
diff --git a/src/bub/config.py b/src/bub/config.py
index c465ff5d..fb427e97 100644
--- a/src/bub/config.py
+++ b/src/bub/config.py
@@ -5,8 +5,6 @@
 from pydantic import Field
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
-from .skills import build_skills_prompt_section
-
 
 class Settings(BaseSettings):
     """Bub application settings."""
@@ -87,16 +85,4 @@ def get_settings(workspace_path: Path | None = None) -> Settings:
     if agents_md_content:
         settings.system_prompt = agents_md_content.strip()
 
-    skills_section = build_skills_prompt_section(workspace_path)
-    if skills_section:
-        settings.system_prompt = _append_prompt_section(settings.system_prompt, skills_section)
-
     return settings
-
-
-def _append_prompt_section(base: str | None, section: str) -> str:
-    base_text = (base or "").strip()
-    section_text = section.strip()
-    if not base_text:
-        return section_text
-    return f"{base_text}\n\n{section_text}"
diff --git a/src/bub/skills/skill-creator/scripts/init_skill.py b/src/bub/skills/skill-creator/scripts/init_skill.py
index f90703ec..f722b98d 100644
--- a/src/bub/skills/skill-creator/scripts/init_skill.py
+++ b/src/bub/skills/skill-creator/scripts/init_skill.py
@@ -18,7 +18,6 @@
 import sys
 from pathlib import Path
 
-from generate_openai_yaml import write_openai_yaml
 
 MAX_SKILL_NAME_LENGTH = 64
 ALLOWED_RESOURCES = {"scripts", "references", "assets"}
@@ -296,15 +295,6 @@ def init_skill(skill_name, path, resources, include_examples, interface_override
         print(f"[ERROR] Error creating SKILL.md: {e}")
         return None
 
-    # Create agents/openai.yaml
-    try:
-        result = write_openai_yaml(skill_dir, skill_name, interface_overrides)
-        if not result:
-            return None
-    except Exception as e:
-        print(f"[ERROR] Error creating agents/openai.yaml: {e}")
-        return None
-
     # Create resource directories if requested
     if resources:
         try:
diff --git a/tests/test_agent_contract.py b/tests/test_agent_contract.py
index 376d821d..850ded9e 100644
--- a/tests/test_agent_contract.py
+++ b/tests/test_agent_contract.py
@@ -4,6 +4,7 @@
 
 import json
 from dataclasses import dataclass
+from pathlib import Path
 from types import SimpleNamespace
 from typing import Any, ClassVar
 
@@ -77,6 +78,15 @@ def _handler(event: ToolEvent) -> None:
     return _handler
 
 
+def _write_skill(root: Path, folder: str, *, name: str, description: str) -> None:
+    skill_dir = root / folder
+    skill_dir.mkdir(parents=True, exist_ok=True)
+    (skill_dir / "SKILL.md").write_text(
+        (f"---\nname: {name}\ndescription: {description}\n---\n\nSkill instructions.\n"),
+        encoding="utf-8",
+    )
+
+
 def test_tool_result_payload_is_structured_json(tmp_path, monkeypatch) -> None:
     monkeypatch.setenv("BUB_MODEL", "openai:gpt-4o-mini")
     monkeypatch.setattr("bub.agent.core.LLM", _FakeLLM)
@@ -163,3 +173,96 @@ def test_agent_allows_completion_with_verification_evidence(tmp_path, monkeypatc
     result = agent.respond([{"role": "user", "content": "please verify completion and then conclude"}])
 
     assert result == "verified completion"
+
+
+def test_agent_refreshes_skills_per_respond(tmp_path, monkeypatch) -> None:
+    monkeypatch.setenv("BUB_MODEL", "openai:gpt-4o-mini")
+    monkeypatch.setenv("HOME", str(tmp_path / "home"))
+
+    captured_system_prompts: list[str] = []
+
+    class _CaptureChat:
+        def raw(self, *, messages: list[dict[str, Any]], tools: list[Any], max_tokens: int | None = None) -> object:
+            _ = (tools, max_tokens)
+            content = messages[0].get("content", "")
+            captured_system_prompts.append(str(content))
+            return _text_response("done")
+
+    class _CaptureLLM:
+        def __init__(self, model: str, api_key: str | None = None, api_base: str | None = None) -> None:
+            _ = (api_key, api_base)
+            provider, model_name = model.split(":", 1)
+            self.provider = provider
+            self.model = model_name
+            self.chat = _CaptureChat()
+            self.tools = _FakeTools(outputs=["ok"])
+
+    monkeypatch.setattr("bub.agent.core.LLM", _CaptureLLM)
+
+    project_skills = tmp_path / ".agent" / "skills"
+    _write_skill(project_skills, "alpha", name="alpha-skill", description="alpha")
+
+    agent = Agent(context=Context(tmp_path), tools=[SimpleNamespace(name="fs_read")])
+    first = agent.respond([{"role": "user", "content": "first"}])
+    assert first == "done"
+
+    _write_skill(project_skills, "beta", name="beta-skill", description="beta")
+    second = agent.respond([{"role": "user", "content": "second"}])
+    assert second == "done"
+
+    assert len(captured_system_prompts) == 2
+    assert "<name>alpha-skill</name>" in captured_system_prompts[0]
+    assert "<name>beta-skill</name>" not in captured_system_prompts[0]
+    assert "<name>beta-skill</name>" in captured_system_prompts[1]
+
+
+def test_agent_refreshes_skills_within_same_respond_loop(tmp_path, monkeypatch) -> None:
+    monkeypatch.setenv("BUB_MODEL", "openai:gpt-4o-mini")
+    monkeypatch.setenv("HOME", str(tmp_path / "home"))
+    project_skills = tmp_path / ".agent" / "skills"
+    _write_skill(project_skills, "alpha", name="alpha-skill", description="alpha")
+
+    captured_system_prompts: list[str] = []
+
+    class _LoopChat:
+        def __init__(self) -> None:
+            self.calls = 0
+
+        def raw(self, *, messages: list[dict[str, Any]], tools: list[Any], max_tokens: int | None = None) -> object:
+            _ = max_tokens
+            content = messages[0].get("content", "")
+            captured_system_prompts.append(str(content))
+            self.calls += 1
+            if tools and self.calls == 1:
+                return _tool_call_response(name="fs_read", arguments='{"path":"foo.txt"}')
+            return _text_response("done")
+
+    class _LoopTools:
+        def __init__(self) -> None:
+            self.calls = 0
+
+        def execute(self, call: dict[str, Any], *, tools: list[Any] | None = None) -> str:
+            _ = (call, tools)
+            if self.calls == 0:
+                _write_skill(project_skills, "beta", name="beta-skill", description="beta")
+            self.calls += 1
+            return "ok"
+
+    class _LoopLLM:
+        def __init__(self, model: str, api_key: str | None = None, api_base: str | None = None) -> None:
+            _ = (api_key, api_base)
+            provider, model_name = model.split(":", 1)
+            self.provider = provider
+            self.model = model_name
+            self.chat = _LoopChat()
+            self.tools = _LoopTools()
+
+    monkeypatch.setattr("bub.agent.core.LLM", _LoopLLM)
+
+    agent = Agent(context=Context(tmp_path), tools=[SimpleNamespace(name="fs_read")])
+    result = agent.respond([{"role": "user", "content": "check"}])
+
+    assert result == "done"
+    assert len(captured_system_prompts) >= 2
+    assert "<name>beta-skill</name>" not in captured_system_prompts[0]
+    assert "<name>beta-skill</name>" in captured_system_prompts[1]
diff --git a/tests/test_bub.py b/tests/test_bub.py
index 95ee8dee..4814feab 100644
--- a/tests/test_bub.py
+++ b/tests/test_bub.py
@@ -22,12 +22,10 @@ def test_settings_with_agents_md(self, tmp_path, monkeypatch):
         agents_md = tmp_path / "AGENTS.md"
         agents_md.write_text("System prompt from AGENTS.md")
         settings = get_settings(tmp_path)
-        assert settings.system_prompt is not None
-        assert settings.system_prompt.startswith("System prompt from AGENTS.md")
-        assert "<available_skills>" in settings.system_prompt
+        assert settings.system_prompt == "System prompt from AGENTS.md"
 
     def test_settings_with_skills_section(self, tmp_path, monkeypatch):
-        """Test settings include available skills metadata."""
+        """Test settings do not inline available skills metadata."""
         monkeypatch.setenv("HOME", str(tmp_path / "home"))
         skill_dir = tmp_path / ".agent" / "skills" / "code_review"
         skill_dir.mkdir(parents=True)
@@ -38,9 +36,7 @@ def test_settings_with_skills_section(self, tmp_path, monkeypatch):
 
         settings = get_settings(tmp_path)
         assert settings.system_prompt is not None
-        assert "<available_skills>" in settings.system_prompt
-        assert "<name>code-review</name>" in settings.system_prompt
-        assert "code_review/SKILL.md" in settings.system_prompt
+        assert "<available_skills>" not in settings.system_prompt
 
 
 class TestTools: