From agent-team
Activate when an agent prompt is being assembled and you need to understand how directory context, file trees, and key-file hints are produced. Governs the lightweight orientation block that ships in every agent prompt.
How this skill is triggered — by the user, by Claude, or both
Slash command
/agent-team:context-loadingThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
This skill describes how `scripts/load-context.sh` builds the orientation block
This skill describes how scripts/load-context.sh builds the orientation block
that goes into every agent prompt. The design favours awareness over volume:
agents see what exists and which files matter, then read those files themselves
with their own tools.
Earlier versions of the dispatcher inlined up to 30 files × 4KB each into every agent prompt — ~120KB of noise. Agents skimmed it, missed details, and used the context window for indexing instead of thinking. The new approach:
This trades a heavier first read step for vastly better signal-to-noise across the rest of the round.
bash ${CLAUDE_PLUGIN_ROOT}/scripts/load-context.sh \
"$PROJECT_ROOT" "$CONTEXT_DIRS_JSON"
Output structure:
## Project Context
_<one-line note explaining tree-only mode>_
### `frontend/`
<file tree, up to 80 lines, sorted>
## Key Files — Read These First (frontend)
- `frontend/package.json`
- `frontend/next.config.ts`
- `frontend/app/page.tsx`
- ...
context_dirsnode_modules/, .git/, __pycache__/, dist/, .next/,
build/, .venv/, venv/, .cache/, coverage/A file is flagged as "read first" if its basename matches one of these orientation-class names:
| Category | Examples |
|---|---|
| Build/manifest | package.json, pyproject.toml, tsconfig.json, next.config.*, tailwind.config.*, postcss.config.mjs |
| Container | Dockerfile, docker-compose.yml |
| Entry points | main.py, app.py, index.ts(x), index.js, server.py, __main__.py |
| Framework entries | page.tsx, layout.tsx |
| Docs | README.md, PLAN.md, ARCHITECTURE.md, CLAUDE.md |
| Schema | schema.sql, schema.prisma |
Up to 10 key files surface per directory. The agent reads them itself before writing any code — that's part of the dispatcher's instructions in the prompt.
The default is never inline. The narrow exception: if a file is shorter than
~20 lines AND it's load-bearing for the task, the coordinator may quote it
inline in the agent's mission brief (the ${TASK} portion of the prompt).
load-context.sh itself never does this — it's the coordinator's call.
dispatch-agent.sh reads the first 50 lines of planning/PLAN.md (if present)
and embeds them above the file tree. The PLAN file is the closest thing to a
project-wide spec, so it's worth always being in the prompt — but only the head
of it. Agents read the rest themselves on demand.
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub usmannoor5/plugins --plugin agent-team