Token-efficient delegation

Distilled from the 2026 token-optimization writeups (BuildToLaunch, GitConnected, affaan-m/everything-claude-code) and the official Anthropic prompt-caching docs. Use this skill in the planning phase: every task in a roadmap should have an explicit executor assignment from the table below, not a default to Opus.

Why this matters

Claude re-reads the entire conversation on every turn. By message 50, you're paying to re-read 49 prior messages before any new work. Most cost overruns aren't bad prompts — they're context bloat plus mismatched model tiers.

Rule 1 — Tier the model to the task

When delegating from inside Claude Code, be explicit: Agent({ subagent_type: "general-purpose", model: "haiku", prompt: "..." }).

Rule 2 — Delegate heavy reads to subagents

Any task that would have me read 3+ files or 15K+ tokens of source goes to an Explore or general-purpose subagent. The subagent's accumulated context dies with it; only the summary lands in main context. Never duplicate work — if a subagent is researching, don't also grep myself.

Rule 3 — Self-contained prompts for external tools

Aider and Gemini Code Assist don't see the planning conversation. When handing off to them:

• Write the prompt as a complete unit: file paths, line numbers, exact change, why it matters
• Reference AGENTS.md for project conventions (both tools import it via .aider.conf.yml / GEMINI.md)
• Don't write "based on the plan" or "based on findings" — paste the actual instruction

Rule 4 — One topic per chat

When switching to an unrelated task, /clear and start fresh. For continuity within a topic, /compact at logical breakpoints (after exploration, after a bug is fixed, before pivoting). Never compact mid-implementation. The /session-handoff skill is the right pattern when a topic genuinely needs to span sessions.

Rule 5 — Specific file paths over folder vagueness

"Edit src/components/SiteHeaderNav.tsx:168-179 to convert the <Link> to a DropdownMenu" costs ~50 input tokens. "Update the nav to add a dropdown" forces Claude to grep, list, and read until it finds the right file — easily 10× the cost.

Rule 6 — Extended thinking budget

For non-reasoning tasks (renames, mechanical edits, content generation), set MAX_THINKING_TOKENS=10000 (or 0). Default is ~32K — most edits don't need it. Toggle thinking before sending the message; toggling mid-conversation invalidates the messages cache.

Rule 7 — Prompt caching hygiene

The Anthropic API caches in the hierarchy tools → system → messages; a change at any level invalidates that level and below. Practical implications inside Claude Code:

• Don't toggle fast mode, web search, or extended thinking mid-task — each toggle invalidates the messages cache
• Long-running agent loops (e.g. enrichment fan-out, scheduled /loop jobs) benefit from 1-hour cache TTL on the system prompt; 5-min default for normal sessions
• Prompt cache 5-min TTL means: sleeping >300s in a /loop blows the cache. Stay <270s for hot work, or commit to ≥1200s for genuinely idle waits — never split the difference

Rule 8 — CLI over MCP when both exist

gh pr view 34 returns the PR JSON. The GitHub MCP returns the same data wrapped in tool-definition tokens that stay in context forever. Same for supabase, aws, vercel. Default to CLI.

Rule 9 — Skills load on demand

Don't list-and-load skills reflexively. Only invoke the ones the current task needs. The Skill tool's system reminder is informational — it's not a checklist.

Rule 10 — .claudeignore and CLAUDE.md hygiene

• .claudeignore should exclude node_modules/, .next/, dist/, build/, *.log, .env*. (Verify the project has this.)
• CLAUDE.md is re-read every turn. Keep it under ~500 tokens; link to AGENTS.md for the full runbook.

Quick checklist before delegating

• Is this mechanical? → Aider or Gemini Code Assist
• Is this single-file with clear spec? → Haiku subagent
• Is this multi-file with judgment? → Sonnet subagent
• Is this architectural / ambiguous? → keep on Opus
• Did I write the prompt as if the executor has zero context? (For external tools, yes.)
• Am I duplicating a subagent's work?

How to apply this in a planning session

When producing a multi-PR roadmap, every PR gets an explicit "Executor:" line. Don't default to "I'll do it" — that means Opus. Examples:

PR 1 — Quick wins. Executor: Gemini Code Assist (in VS Code), one edit at a time. Each change is single-file, single-purpose; perfect for free-tier sidebar edits with the user driving.

PR 6 — Schema rework. Executor: Opus (this session) for the design + migration; Sonnet subagent for the route restructuring; Gemini for individual file moves.

The user can then disagree with the assignment ("no, do PR 4 yourself"), but the default surfaces the cheap path.