메뉴
Every agent loads this skill at the start of every session to consult the long-term knowledge base before rediscovering prior work. The primary agent may delegate global KB writes; subagents have read access and report candidate learnings upward.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | kb |
| description | Every agent loads this skill at the start of every session to consult the long-term knowledge base before rediscovering prior work. The primary agent may delegate global KB writes; subagents have read access and report candidate learnings upward. |
The knowledge base (KB) is your long-term memory. It is a collection of HTML
files at ~/.config/opencode/kb/. Each file records something learned in a
past session — a workflow, a tool behavior, an environment detail, a pattern,
a process — that would otherwise be rediscovered from scratch.
This is not documentation for the user. It is memory for you.
Every agent must load this skill at the start of every session. Read the KB directory listing and build a lightweight mental index:
~/.config/opencode/kb/
For each .html file found, extract from its <head>:
<title> — the entry name<meta name="description"> — a one-line summary<meta name="tags"> — comma-separated topics<meta name="priority"> — high, normal, or low<meta name="updated"> — last updated dateHold this index in mind for the session. Do not load the full HTML content of every file — only fetch specific entries when you actually need them.
Check the updated date of each entry. If any entry has not been updated in
180 days or more, handle the observation according to your role:
"N KB entries haven't been updated in 6+ months: [title1], [title2], ... Worth reviewing when you get a chance."
State or report this once, do not repeat it. Do not delete or alter stale entries merely because of their age.
Before attempting something you might have done before — using a tool, running a workflow, looking something up, interacting with a service — pause and check the index. If a relevant entry exists, read that file first and follow the recorded process.
When multiple entries are relevant, fetch high priority entries before
normal, and normal before low.
When you use a KB entry during work and observe that the recorded behavior does not match reality — a command fails, a path is wrong, or a process has changed — handle it according to your role:
kb-writer to update the entry with the correct
information, noting what changed and why.kb-writer or announce an update to the user.
Report the outdated behavior, observed replacement, evidence, and affected
entry to the primary agent.Do not silently leave incorrect entries in place.
When you discover something non-trivial that would be useful to remember in future sessions, first distinguish project-specific information from durable, cross-project knowledge:
context-capture.kb-writer.Then handle the candidate according to your role:
context-capture with
target: agents_md for broad project facts and conventions, or
target: skill for a focused repeatable workflow. Pass the absolute
project_root and concise content with useful supporting context. For
target: skill, also pass skill_name and skill_description.kb-writer with:
high (frequently needed or critical), normal (default),
or low (edge case or rarely needed)kb-writer or context-capture. Report candidate
global and project-specific learnings to the primary agent with supporting
context so it can classify and route them.The primary agent and ordinary subagents never edit KB files directly. Only the
primary agent delegates creates or updates, and the explicitly delegated
kb-writer is the authorized writer.
Each entry is a valid HTML file. Filename: <slug>.html where the slug is
lowercase, hyphen-separated. Examples: gitlab-mr-diff.html, jira-ticket-lookup.html.
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<title>[Concise title]</title>
<meta name="description" content="[One-line summary]">
<meta name="tags" content="[tag1, tag2, tag3]">
<meta name="priority" content="[high|normal|low]">
<meta name="created" content="YYYY-MM-DD">
<meta name="updated" content="YYYY-MM-DD">
</head>
<body>
<h1>[Title]</h1>
<p>[Context: why this matters, when to use it]</p>
<h2>Steps</h2>
<ol>
<li>...</li>
</ol>
<h2>Examples</h2>
<pre><code>[commands or code]</code></pre>
<h2>Notes</h2>
<ul>
<li>[Caveats, gotchas, environment-specific details]</li>
</ul>
</body>
</html>
Sections are optional — include what is relevant. Keep entries factual and concise. These are reference notes, not tutorials.