Skip to content

SKILL.md

Latest commit

 

History

History
229 lines (175 loc) · 9.58 KB

SKILL.md

File metadata and controls

229 lines (175 loc) · 9.58 KB
Raw
name delegate
description Unified external AI delegation — route tasks to Codex, Gemini, or Grok based on cost/quality tradeoffs. Use when user says 'delegate', 'codex review', 'ask gemini', 'ask grok', 'second opinion', or needs external model.

Delegate — Unified External AI Delegation

Route tasks to the best external AI tool: Codex CLI, Gemini CLI, or Grok CLI via their native binaries.

Prerequisites: codex, gemini, and grok already installed and authenticated locally.

Routing Matrix

Task Tool Command Effort
Code review (uncommitted) Codex codex exec on git diff medium
Code review (vs branch) Codex codex exec on git diff <branch> medium
Deep code review / tech debt Codex codex exec xhigh
Architecture / design review Codex codex exec xhigh
Refactoring strategy Codex codex exec xhigh
Tech debt analysis Codex codex exec xhigh
Security audit Codex codex exec high
Second opinion on my code Codex codex exec medium
Plan validation Codex codex exec xhigh
Bug investigation Codex codex exec xhigh
Large codebase analysis Gemini gemini --all-files —
Full project scan / patterns Gemini gemini --all-files —
Deep architecture analysis Gemini gemini --all-files -m gemini-3.1-pro-preview —
File review (>200 lines) Gemini cat FILE | gemini -p —
Documentation / summary Gemini gemini -p —
Rust code review Codex codex exec high
Quick second opinion (free) Grok grok -p —
Web-aware code question Grok grok -p —
Third independent opinion Grok grok -p —

When multiple tools fit: prefer cheapest → Gemini (free tier) / Grok (free, grok.com) > Codex (paid). When quality matters most: GPT-5.4 (xhigh) > Codex 5.3 (xhigh) > Gemini 3.1 Pro. Rust-specific: Codex is best for ownership, lifetimes, concurrency reasoning. Grok niche: free via grok.com, agentic, built-in web search (Codex has none) — good for web-aware questions and a cheap third opinion.

1. Codex CLI

Non-interactive execution

# Read-only task
codex exec \
  -c model_reasoning_effort="EFFORT" \
  --sandbox read-only \
  --full-auto \
  --skip-git-repo-check \
  "PROMPT"

# Write-capable task (Codex modifies files)
codex exec \
  -c model_reasoning_effort="EFFORT" \
  --sandbox workspace-write \
  --full-auto \
  "PROMPT"

Flags reference:

  • -c model_reasoning_effort="..." — minimal | low | medium | high | xhigh
  • --sandbox read-only|workspace-write — control file modification scope
  • --full-auto — auto-accept tool calls (non-interactive)
  • --skip-git-repo-check — skip the safety check when running outside a git repo

Reasoning effort auto-select:

Task complexity Effort
Trivial (format, explain) low
Standard (review, check) medium
Complex (debug, investigate) high
Critical (architecture, security) xhigh

Code review against branch

# Pipe diff into Codex
git diff main..HEAD | codex exec \
  -c model_reasoning_effort="medium" \
  --sandbox read-only \
  --full-auto \
  --skip-git-repo-check \
  "Review this diff for correctness, security, and maintainability issues. Cite file:line."

Codex prompt template

Structure prompts for Codex with mandatory goal articulation before execution:

TASK: [One sentence, atomic goal]
EXPECTED OUTCOME: [What the output should look like]
CONTEXT: [Relevant code, state, background — paste snippets]
CONSTRAINTS: [Technical limits, patterns to follow]
MUST DO: [Non-negotiable requirements]
MUST NOT DO: [Anti-patterns to avoid]
OUTPUT FORMAT: [How to structure the response]

Before executing: restate the goal in your own words. Identify implicit constraints and assumptions. Only then proceed.

For simple tasks, just TASK + CONTEXT is enough. Scale up sections with complexity. The "Before executing" line is always included — even in minimal prompts.

2. Gemini CLI

Headless mode

# Simple prompt
gemini -p "PROMPT" -m gemini-3-flash-preview -e none -o text

# Full project analysis (1M context window)
gemini --all-files -p "PROMPT" -m gemini-3-flash-preview -e none

# Pipe file content
cat src/auth.py | gemini -p "Review for security issues" -m gemini-3-flash-preview -e none

Flags reference:

  • -p "PROMPT" — required for headless (non-interactive mode)
  • -m MODEL — model selection
  • -e none — always use — disables MCP extensions, saves 2-3s startup
  • -o text — output format: text (default), json, stream-json
  • --all-files or -a — include ALL project files in context (uses 1M window)
  • --include-directories DIR1,DIR2 — scope to specific dirs
  • -y / --yolo — auto-accept all actions
  • --approval-mode plan — read-only mode (safest)

Model selection:

Model Use for Context Speed
gemini-3-flash-preview Default — most tasks, fast 1M Fast
gemini-3.1-pro-preview Deep analysis, complex reasoning 1M Slower
gemini-3.1-flash-lite-preview Budget tasks, high volume 1M Fastest

Gemini prompt style: Keep prompts direct and specific. Gemini works best with a clear task statement, specific things to look for, and desired output structure. Avoid verbose 7-section format — Gemini prefers concise instructions.

Goal articulation for Gemini: Append to every prompt: First, restate the goal and key constraints in one sentence, then proceed. This is lightweight enough for Gemini's style while forcing structured reasoning.

3. Grok CLI

Agentic coding CLI (xAI), free via grok.com login. Full tool use and built-in web search — the only CLI here with native web access. Best as a fast free third opinion or for web-aware questions.

Headless mode

# Read-only (review / second opinion — no shell, edit, or web)
grok -p "PROMPT" --tools "read_file,grep,list_dir" --output-format plain

# Web-aware task (web_search / web_fetch on by default)
grok -p "PROMPT" --output-format plain

# JSON output for parsing
grok -p "PROMPT" --output-format json

Flags reference:

  • -p "PROMPT" — required for headless (alias --single)
  • -m MODEL — grok-build (default) or grok-composer-2.5-fast
  • --tools "read_file,grep,list_dir" — allowlist; keeps the run read-only
  • --disallowed-tools "..." — denylist (e.g. run_terminal_cmd,search_replace,web_search,web_fetch)
  • --allow RULE / --deny RULE — ToolPrefix(glob) permission gates (e.g. --deny "Bash(rm*)")
  • --output-format plain|json|streaming-json — plain default
  • --always-approve — auto-approve tool executions (write tasks)
  • -s ID / -r ID / -c — named session / resume / continue

⛔ No --effort — grok-build rejects reasoningEffort with HTTP 400; the model has no effort knob.

Models:

Model Use for
grok-build Default — agentic coding, review, second opinion
grok-composer-2.5-fast Lighter / faster tasks

Grok prompt style: accepts structured prompts like Codex. Keep a Before executing: restate the goal… line. For review / second-opinion always pass --tools "read_file,grep,list_dir" to prevent edits, shell, and web. Single grok.com account — grok models checks login; grok login is interactive.

4. Model Selection Guide

GPT-5.3 Codex and Opus 4.6 have converged in capability. GPT-5.4 stands apart for deep work.

Task Type Model Effort Why
Planning & architecture GPT-5.4 (Codex) xhigh Most thorough, raises corner cases, prevents arch drift
Refactoring strategy GPT-5.4 (Codex) xhigh Best at minimizing tech debt
Bug investigation (complex) GPT-5.4 (Codex) xhigh Superior context gathering, finds non-trivial root causes
Verification & review GPT-5.4 (Codex) xhigh Catches what other models miss, "pedantic" but correct
Implementation of plans GPT-5.3 Codex high-xhigh Fast execution, good at following instructions
Quick code review GPT-5.3 Codex medium Fast turnaround for standard reviews
Interactive fast work GPT-5.3 Codex medium Best speed/quality for conversational coding
Documentation / summary Gemini Flash — Cheap, fast, 1M context
Whole-project pattern scan Gemini Flash (--all-files) — 1M window beats everything for breadth

GPT-5.4 traits to expect:

  • Slow start (10-20 min context gathering) — this is a feature, not a bug
  • Will flag corner cases and impossibilities — feels "pedantic" but catches real issues
  • More agentic — can autonomously drive multi-step tasks to completion
  • Best at preventing architectural drift and reducing tech debt

When NOT to use GPT-5.4:

  • Quick questions or simple tasks (use 5.3 Codex low)
  • Time-sensitive interactive work (use 5.3 Codex)
  • Tasks that don't benefit from deep reasoning (use Gemini flash)

5. Error Handling

  • Codex timeout (>60s) — retry with lower -c model_reasoning_effort=...
  • Codex auth issues — run codex login (or your install's documented login flow)
  • Gemini fails — try without -e none, or different model
  • Gemini opens browser OAuth in headless — stop and ask user to authenticate first
  • Grok 400 … does not support parameter reasoningEffort — remove --effort (grok-build has no effort knob)
  • Grok asks to log in / hangs on OAuth — run grok login, then retry