Skip to content

electronx/wiredkits-free / skills/pick-model-effort

pick-model-effort

Unverified

Choose the Claude model and effort level for a task or a subagent, so cheap work runs on cheap models and hard work gets the strong ones. Use before spawning subagents or parallel agents, when planning a multi-step task, when asked "which model should I use", or when setting up agents in .claude/agents.

Pick the model and effort

The goal is accurate work without paying top-model prices for searching files. Pick the settings yourself, say what you picked when you delegate, and move on.

1. Sort the work

Split the task into pieces and give each one a class:

ClassExamplesModelEffort
Lookupfind files, grep, read docs, list callers, summarize logs, classify, reformathaikulow
Routine buildclear-scope edits, tests from a spec, renames, small features, docssonnetmedium
Judgmentcode review, bug hunting, hard debugging, security, migrationsopushigh
Designarchitecture, plans, research with trade-offs, decisions that are costly to undoopusxhigh
Hardestlong autonomous runs, or a problem where opus already fell shortfablehigh or xhigh

Rules of thumb:

  • When unsure, go higher for decisions and lower for mechanical work.
  • A stronger model at lower effort usually beats a weaker model at high effort.
  • max effort only when being right matters far more than cost or time.
  • Lookups are where most tokens go. Moving them off the main model is the biggest saving.
  • Use only models your account offers. If fable isn't available, use opus at xhigh.

2. Know how Claude Code picks a subagent's model

First match wins:

  1. model passed for that one call (the Agent tool's model parameter)
  2. model: in the agent file (inherit means the main session's model)
  3. The CLAUDE_CODE_SUBAGENT_MODEL environment variable
  4. The main session's model

Effort comes from effort: in the agent file or the call. The CLAUDE_CODE_EFFORT_LEVEL environment variable overrides both.

The trap: the built-in Explore, Plan and general-purpose agents inherit the main model. Six parallel Explore agents in an Opus session are six Opus agents. Pass model: "haiku" when you spawn them for lookups, or use the searcher agent below.

3. Use the agents in this kit

This kit installs four agents with sensible defaults:

AgentModel, effortToolsUse for
searcherhaiku, lowread-onlyfinding code, lookups, reading many files and returning a short summary
implementersonnet, mediumallroutine edits and tests with a clear spec
revieweropus, highread-onlyreviewing a diff, hunting bugs, checking a plan
architectopus, xhighread-onlydesign, plans, trade-offs

If you need a different setting for one call, pass model (and effort, where your version supports it) on that call instead of editing the agent.

If an agent isn't installed (for example, the skill was copied without the agents), pass model on each Agent call using the table in step 1.

4. When you delegate

  • In one line, say which agent or model and effort you chose and why, for example: "Searching with searcher (haiku), then reviewer (opus, high) on the diff."
  • Send heavy reading to a subagent and ask for a short summary with file paths, not a dump. That keeps the main context small, which saves tokens on every later turn.
  • Run independent lookups in parallel, on the cheap model.
  • Don't give a subagent a decision it can't see the context for. Keep the decision in the main session and delegate the gathering.

5. Keep the main session cheap

  • If the next task is unrelated to the last one, suggest /clear or a new session. A long context makes every turn cost more.
  • If the user asks for something that breaks these rules (an Opus or Fable subagent at high effort for a file search, say), point it out in one line with the cheaper option, then do what they decide. Don't block them or repeat the point.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 467kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 319kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 319kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 319kUpdated Claude CodeCodexVerifiedInstall