Effort Selector

---
name: effort-selector
description: Set the AI's thinking depth — low, medium, high, xhigh, or max effort — matched to how hard the task is, not how important it feels. Use when asked what effort level to use, whether to use extended thinking, why the AI overthinks simple jobs or under-thinks hard ones, how to make responses faster or cheaper, tune an agent's reasoning depth, or someone says 'it's thinking forever on a caption'. For which model, see model-picker. For session cost, see token-budgeter.
metadata:
  version: 1.0.0
---

# Effort selector

Separate from the model, effort controls **how much the AI thinks before it
answers**. More thinking means better answers on hard problems — and more tokens,
more time, more cost. It is one of your biggest token levers, and it is not the
same lever as the model.

The rule that prevents both failure modes: **match effort to how hard the task
is, not to how important it feels.** Importance tempts people to max; difficulty
is what actually earns it.

## Before you start

| Input | Why |
| --- | --- |
| The task, concretely | Difficulty is a property of the task, not the client |
| Whether it is multi-step/agentic | Long autonomous chains are what xhigh exists for |
| The failure cost of a wrong answer | max is reserved for cannot-be-wrong, once |

## Step 1 — Know the five levels

| Level | Thinking | Use for |
| --- | --- | --- |
| **low** | Least, fewest tokens | Captions, formatting, simple classification, cheap sub-agents |
| **medium** | A balance | Everyday drafting where speed matters and quality is decent |
| **high** | **The default, usually the sweet spot** | Most real work: blog posts, ad copy, most agent runs — best quality-to-token balance |
| **xhigh** | More thinking, more tokens | The hardest agentic, multi-step work: a research agent, a complex build. Not a caption |
| **max** | The most | When correctness matters more than cost. Your single highest-stakes task. Can overthink simple jobs |

## Step 2 — Apply the marketer's rule of thumb

**Start every agent at high.** Drop to low/medium for the light, repetitive parts
(formatting, bulk captions). Push to xhigh only for genuinely hard multi-step jobs
like research or a full audit. Save max for the one thing that absolutely cannot
be wrong.

| Task | Effort | Why |
| --- | --- | --- |
| Bulk ad captions, file cleanup | low | Simple and high-volume; save the tokens |
| LinkedIn post, carousel slide | medium | Fast, decent quality, cheap |
| Blog post, ad copy, SEO metadata | high | The everyday sweet spot |
| Research agent, competitor audit | xhigh | Long, multi-step, needs deep thinking |
| A pillar post that must be flawless | max | Correctness over cost, this once |

## Step 3 — Cross it with the model, don't confuse it

Model picks the engine; effort picks how hard it revs. They compound:

- Haiku + low = the bulk cell. Sonnet + high = the daily workhorse.
- Opus + xhigh = the audit. Fable + xhigh = the overnight research run.
- Opus + max = the one flagship artifact, once.

Raising effort on a too-small model does not buy the missing capability, and a
premium model at low effort wastes what you paid for. Mismatches are the silent
budget leak.

## Output

```
# Effort plan: <task or agent>

Default: high (the sweet spot)
| Step | Model × Effort | Why |
| bulk captions | Haiku × low | volume, simple |
| draft posts | Sonnet × high | everyday real work |
| deep audit | Opus × xhigh | multi-step reasoning |
| flagship pillar | Opus × max | cannot be wrong, once |

max used: <0 or 1 places> — more than one means importance-inflation
```

## When it breaks

| What you see | What it means | The fix |
| --- | --- | --- |
| Minutes of thinking on a caption | Effort set by importance, not difficulty | Drop to low/medium; difficulty earns depth |
| Hard multi-step job comes back shallow | high on xhigh-shaped work | Push to xhigh for the research/audit class |
| Everything runs at max "to be safe" | Importance inflation | max is one task, once; the table reassigns the rest |
| Fast but sloppy everywhere | Defaulted low to save cost | high is the sweet spot for real work — the saving was false |
| Premium model, low effort | Levers conflated | Engine and depth compound; match both to the task |
| Costs crept up with no quality gain | xhigh became the quiet default | Reset the default to high; xhigh by exception |

Never set effort by how nervous the stakeholder is. Importance inflates effort;
only difficulty earns it — a flawless simple task at max is just a slow, expensive
simple task.

## Rules

- **high is the default**, because it is the tested quality-to-token sweet spot
  for real work.
- **Difficulty earns depth, importance does not**, because max on easy work
  overthinks and overbills without improving anything.
- **max appears at most once in a plan**, because "highest-stakes task" is by
  definition singular.
- **Set effort per step, not per agent**, because one pipeline contains captions
  and audits, and they deserve different depths.
- **Effort and model are separate levers**, because conflating them either wastes
  a premium engine or overdrives a small one.

## Related skills

- **model-picker** — the engine this lever revs; assign both per step.
- **token-budgeter** — effort is one of the biggest token levers in the 13 moves.
- **claude-md-writer** — record the effort defaults in the agent's driver file.