Gemini Smart Delegation Skill

Route tasks to Google's Gemini CLI (1M token context window) intelligently. The key behavior: always evaluate FIRST, then decide — never start a big task without checking.

Step 1: Read the Experience Log

Before doing anything else, read the experience log:

cat ~/.claude/skills/gemini/experience.md

The log contains:

• [AUTO_APPROVE] entries — task types to always send to Gemini, no prompt needed
• [AVOID] entries — task types where Gemini underperformed, handle with Claude instead

Step 2: Score the Task

Strong Gemini signals (+2 each)

• Files involved are likely >50KB total (codebase scan, large logs, big documents)
• Task is purely read/analysis — no file editing needed
• Task matches an [AUTO_APPROVE] pattern in experience.md
• User explicitly says "use gemini", "ask gemini", "let gemini handle"

Mild Gemini signals (+1 each)

• Multiple files need to be read and cross-referenced
• Task involves summarization or extraction from long content
• Token conservation would meaningfully benefit the conversation

Disqualifiers (skip Gemini entirely)

• Task matches an [AVOID] pattern in experience.md → handle with Claude, mention why
• Task requires writing or editing files
• Task requires iterative tool-use loops
• Task is short and simple

Step 3: Decision Branch

Score >= 2 AND matches an [AUTO_APPROVE] pattern:

→ Go straight to Gemini — skip the question → Briefly note: "Routing to Gemini (auto-approved: [pattern name])"

Score >= 2 but NO prior auto-approve for this pattern:

→ Use AskUserQuestion with this exact structure:

Question: "This task looks like a good fit for Gemini (large context / token savings). How should I handle it?"
Header: "Route to Gemini?"
Options:
  1. label: "Use Gemini & remember this task type"
     description: "Delegate to Gemini now, and auto-route similar tasks in the future without asking"
  2. label: "Use Gemini, just this once"
     description: "Delegate to Gemini now, but ask again next time"
  3. label: "Claude handles it"
     description: "Skip Gemini for this task"

Score < 2:

→ Handle with Claude normally, no question needed

Step 4: After the User Chooses

Option 1 — "Use Gemini & remember this task type":

1. Append an [AUTO_APPROVE] entry to experience.md (format below)
2. Run Gemini
3. Present output labeled [Gemini Response]

Option 2 — "Use Gemini, just this once":

1. Run Gemini
2. Present output labeled [Gemini Response]
3. Do NOT write to experience.md

Option 3 — "Claude handles it":

1. Handle with Claude normally
2. Do NOT write to experience.md

Step 5: Running Gemini

Standard headless call

gemini -p "COMPLETE SELF-CONTAINED PROMPT" --output-format text --yolo 2>&1

Pipe large file content

cat /path/to/large/file | gemini -p "YOUR INSTRUCTION" --output-format text --yolo 2>&1

Multi-file analysis

gemini -p "$(printf 'Analyze these files:\n\n=== file1 ===\n'; cat file1; printf '\n\n=== file2 ===\n'; cat file2; printf '\n\nTask: YOUR INSTRUCTION')" --output-format text --yolo 2>&1

Important: Gemini headless is single-shot — craft a complete, self-contained prompt with all needed context included.

Step 6: Post-Result Dissatisfaction Detection

After presenting Gemini output, watch the user's next message for dissatisfaction signals:

Signals: "wrong", "not good", "that's off", "redo this", user corrects the output, user asks Claude to redo the same task

When detected:

1. Use AskUserQuestion:

   Question: "Should I mark this task type as a poor fit for Gemini and handle it with Claude going forward?"
   Options:
     - "Yes, log it" — description: "Record this as an [AVOID] pattern"
     - "No, just this once" — description: "Don't log anything"
   

2. If user picks "Yes, log it" → append an [AVOID] entry to experience.md
3. If this task type had an existing [AUTO_APPROVE] entry, update it to [AVOID]
4. Confirm: "Noted. I'll handle this type of task myself from now on."

Experience Log Entry Formats

AUTO_APPROVE entry

---
### [AUTO_APPROVE] YYYY-MM-DD
**Pattern**: one-line description of the task type (e.g., "Full Go codebase read-only analysis")
**Trigger keywords**: keywords that identify this pattern
**Reason**: why Gemini is well-suited for this

AVOID entry

---
### [AVOID] YYYY-MM-DD
**Pattern**: one-line description of the task type
**What went wrong**: brief failure description from user feedback
**Lesson**: what to do instead next time

Decision Table

Source: audi0417/claude-gemini-router — distributed by TomeVault.