Skip to content
Agent Orchestration
Command

/prime

Prime context for fresh session: /prime [task hint] [--quick] [--verbose]

BOOST
From plugin
oh-my-claude
1741 skill7 agents1 command2 MCP
Install
> /plugin marketplace add TechDufus/oh-my-claude
> /plugin install oh-my-claude@oh-my-claude

How it fires

How this command gets triggered: by you, by Claude, or both.

  • Fires itselfClaude auto-loads it when your prompt matches the work.
  • You can call itInvoke it directly when you want it.
  • Slash command/prime

Context preview

What this command does when you run it.

Prime context for fresh session: /prime [task hint] [--quick] [--verbose]

Command definition

prime.md
description: "Prime context for fresh session: /prime [task hint] [--quick] [--verbose]"
argument-hint: "[task hint] [--quick] [--verbose]"
allowed-tools:
  - Bash(git:*)
  - Bash(gh:*)
  - Agent
  - Read
  - Grep
  - Glob

/prime - Context Priming for Fresh Sessions

Rapidly build complete understanding of current work state after `/clear` or new session start. Uses parallel subagents to maximize context gathering while minimizing main context token usage.

Usage

/prime                                    # Full context priming (default)
/prime <task-hint>                        # Prime with upcoming task focus
/prime --quick                            # Essential git state only (faster)
/prime --verbose                          # Include full diffs in report

Examples

/prime                                    # General orientation
/prime migrating postgres init scripts    # Focus exploration on postgres/init
/prime reviewing PR #42                   # Pull PR context to forefront

Arguments: $ARGUMENTS

---

Philosophy

**Problem**: Fresh sessions lack context about in-progress work.

**Solution**: Subagents explore repository state in parallel, then synthesize into a concise briefing. Main context receives only the summary, not raw exploration data.

**Task-Focused**: When a task hint is provided, subagents prioritize context relevant to that task (pyramid approach - start with intent).

---

Phase -1: Parse Arguments

Parse `$ARGUMENTS` to extract:

  • **Task hint**: Any non-flag text (e.g., "migrating postgres init scripts")
  • **Flags**: `--quick`, `--verbose`

---

Phase 0: Quick Snapshot (Main Agent)

Run these commands in parallel via Bash tool:

git branch --show-current
git rev-parse --abbrev-ref HEAD@{upstream} 2>/dev/null || echo "no upstream"
git status --short --branch
git log --oneline -5
git stash list --format="%gd: %s"

Display immediately:

## Quick Snapshot
- **Branch**: <current-branch> (tracking: <upstream-or-none>)
- **Status**: <N files changed, M staged, K untracked>
- **Recent commits**: <last 3-5 one-liners>
- **Stashes**: <count or "none">

If `--quick` flag provided, STOP HERE.

---

Subagent Prompt Template

All subagents receive this wrapper structure. Replace `{AGENT_NAME}`, `{WORD_LIMIT}`, `{COMMANDS}`, `{REQUIREMENTS}`, and `{XML_FIELDS}`:

Agent(
  subagent_type: "general-purpose",
  description: "{description}",
  prompt: """
  <task_context>
  TASK HINT: {task_hint or "None provided - do general analysis"}
  </task_context>

  {AGENT_SPECIFIC_INSTRUCTIONS}

  Provide a CONCISE summary (under {WORD_LIMIT} words).
  If task hint provided, prioritize findings relevant to that task.

  <commands>
  {COMMANDS}
  </commands>

  <analysis_requirements>
  {REQUIREMENTS}
  If task hint provided, add: **Task relevance**: Which findings relate to the hinted task?
  </analysis_requirements>

  <output_format>
  Return structured XML:
  <{agent}_analysis>
    {XML_FIELDS}
    <task_relevance>...</task_relevance>  <!-- only if task hint provided -->
  </{agent}_analysis>
  </output_format>
  """
)

---

Phase 1: Parallel Subagent Exploration

Launch THREE subagents simultaneously (all in single message):

Subagent 1: Git Changes Analyst

**Word limit**: 500 | **Description**: "Analyze git changes"

**Commands**: 1. `git log main..HEAD --format='%h %s' 2>/dev/null || git log -10 --format='%h %s'` 2. `git diff --cached --stat` (staged) 3. `git diff --stat` (unstaged) 4. `git diff --cached` (full staged - summarize, don't echo) 5. `git diff` (full unstaged - summarize, don't echo) 6. `git stash show -p stash@{0} 2>/dev/null` (if stashes exist) 7. `git ls-files --others --exclude-standard | head -20` (untracked)

**Requirements**:

  • **Branch commits**: Group by feature/fix/refactor
  • **Staged/Unstaged/Untracked changes**: Categorize by src/test/config/docs
  • **Stash contents**: Any paused work?
  • **Work narrative**: 2-3 sentences on current developer activity

**XML Fields**: `branch_commits`, `staged`, `unstaged`, `untracked`, `stashes`, `narrative`

Subagent 2: GitHub Context Analyst

**Word limit**: 400 | **Description**: "Analyze GitHub context"

**Commands**: 1. `gh pr list --state open --author @me --json number,title,headRefName,baseRefName,isDraft,updatedAt,reviewDecision --limit 5` 2. `gh pr view --json number,title,body,state,commits,reviews,statusCheckRollup 2>/dev/null` 3. `gh issue list --assignee @me --state open --json number,title,labels,updatedAt --limit 10` 4. `gh run list --limit 3 --json conclusion,name,updatedAt,headBranch`

**Requirements**:

  • If current branch has PR: Draft status? Review feedback? CI status?
  • If branch name contains issue number: Fetch issue, extract requirements
  • If task hint mentions PR/issue: Fetch that directly

**XML Fields**: `current_pr`, `related_issue`, `ci_status`, `other_prs`, `open_issues`

Return `<github_analysis><none>No GitHub context available</none></github_analysis>` if gh CLI fails.

Subagent 3: Project Context Analyst

**Word limit**: 300 | **Description**: "Analyze project context"

**Standard Checks**: 1. `CLAUDE.md` - Look for "## Active Work" section 2. `.claude/CLAUDE.md` - Project-specific instructions 3. `TODO.md` or `TODO` - Tracked tasks 4. Claude Code `/memory` - relevant persisted learnings, decisions, and context 5. `git diff --name-only HEAD~5..HEAD 2>/dev/null | head -20`

**Task-Focused Exploration** (if hint provided):

  • Glob: Find files matching task keywords
  • Grep: Search for relevant patterns
  • Read: Top 3-5 matching files, summarize structure
  • Note file locations for main agent reference

**Requirements**:

  • Active work description and phase from CLAUDE.md
  • Key learnings, decisions, and context from `/memory`
  • Primary directories and patterns from recent changes
  • Relevant files and conventions (if task hint provided)

**XML Fields**: `active_work`, `memory_context` (with learnings/decisions/cont

Read more
Ships withoh-my-claude

Add ultrawork to any prompt for maximum parallel execution

Get the whole plugin