Model Picker

---
name: model-picker
description: Choose the right Claude model per task and per agent — Haiku, Sonnet, Opus, or Fable — so premium thinking is spent only where it shows in the output. Use when asked which model to use, whether to use Opus or Sonnet, why AI costs are high, which model for an SEO or ads or research agent, what model for bulk work, or someone says 'should I just use the best model for everything'. For session cost hygiene, see token-budgeter. For thinking depth, see effort-selector.
metadata:
  version: 1.0.0
---

# Model picker

The model you choose is a cost decision, exactly like choosing a premium placement
versus a cheap one. **Do not run the expensive one for a job the cheap one
handles** — and do not book prime-time TV to sell one extra t-shirt.

The rule that pays for itself: *burn tokens on purpose.* Deconstruct and write on
a strong model; run bulk renaming on a cheap one. Switch any time with `/model`.

## Before you start

| Input | Why |
| --- | --- |
| The task, or the agent being built | Assignment is per job, not per person |
| Where quality shows in the output | The only step that earns premium is the one where the tier is visible |
| Volume — one artifact or hundreds | Volume work goes to the cheap tier by default |

## Step 1 — Know the four engines

| Model | Tier | Best for |
| --- | --- | --- |
| **Haiku** | Fast & cheap | Renaming/formatting files, sorting keyword lists, yes/no classification, grunt-work sub-agents in a bulk run |
| **Sonnet** | Everyday workhorse | LinkedIn posts and carousels, blog drafts, SEO metadata, first-pass ad copy, running the bulk ads pipeline |
| **Opus** | Heavy thinker | The pillar post, competitor analysis, campaign strategy, untangling a messy ad account |
| **Fable** | Autonomous specialist | A research agent reading 30 sources, building a whole content cluster, long overnight runs, your single hardest problem |

Start on the balanced model. Step up only when it was clearly not enough.

## Step 2 — Assign per agent (the house defaults)

| Agent | Default | Why |
| --- | --- | --- |
| SEO agent | Sonnet, Opus for the pillar post | Writing-heavy; "content writing is the only step where the model tier shows up in the output" |
| LinkedIn agent | Sonnet | Short-form, fast, high volume |
| Carousel agent | Sonnet | Structured slides, no deep reasoning |
| Bulk ads creative | Sonnet orchestrates, Haiku for bulk sub-agents | Put the cheap model on the volume work |
| Research agent | Fable or Opus | Long, multi-source, runs on its own |
| Ads agent | Opus | Strategy plus numbers; reasoning matters |
| Competitor analysis | Opus, Fable for a deep audit | Deep reasoning across a lot of input |

## Step 3 — Split within the pipeline

One agent, several models. The SEO pipeline's own split: **write on the strongest
model, publish and verify on a cheaper one** — metadata, publishing calls, and
link patches run Sonnet; file renaming and sitemap pulls run Haiku.

Ask of every step: *does the model tier show in this step's output?* If a reader
could not tell, the step runs cheap.

## Step 4 — Map to other platforms when asked

The same ladder applies to any AI: the cheap/fast tier maps to Haiku, the balanced
everyday tier to Sonnet, the top reasoning tier to Opus/Fable. Pick the cheapest
tier that does the job well, whatever the platform.

## Output

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

| Step | Model | Why the tier shows (or doesn't) |
| deconstruct reference | Opus | judgement visible in the spec |
| write 100 variants | Haiku sub-agents | volume, template-bound |
| orchestrate + QC | Sonnet | balanced judgement, medium stakes |

Default: <model> · Step-ups: <which step, why> · Switch: /model
Premium share target: small — premium is the exception, not the baseline
```

## When it breaks

| What you see | What it means | The fix |
| --- | --- | --- |
| Costs high, output ordinary | Premium model on cheap work | Re-split; premium only where the tier shows |
| Bulk job crawls and overspends | Volume work on a thinking model | Haiku for the repetitive cells; Sonnet orchestrates |
| Quality complaints on the flagship piece | The one step that earned Opus ran on Sonnet | Step up for that step alone, not the whole pipeline |
| "Just use the best for everything" | Cost intuition missing | The ad-account frame: premium placement for one t-shirt |
| Long autonomous run wanders | Multi-step job under-powered | Research-agent-shaped work is Fable/Opus territory |
| Model debates every session | Decision not written down | Put the assignment in the agent's instruction file once |

Never fix a quality problem by upgrading every step's model. Find the one step
where the tier is visible in the output and upgrade that step only.

## Rules

- **Start balanced, step up on evidence**, because premium-by-default is the
  most common silent overspend.
- **Volume work runs cheap**, because a hundred template-bound cells cannot show
  a premium tier's judgement.
- **Split within the pipeline**, because "one agent" is not one job — write and
  verify deserve different engines.
- **Write the assignment into the agent's driver file**, because a decision made
  every session is a decision paid for every session.
- **Model ≠ effort** — they are separate levers; see effort-selector, because
  conflating them doubles the overspend.

## Related skills

- **effort-selector** — the thinking-depth lever on top of the model choice.
- **token-budgeter** — session economics around whichever engine runs.
- **claude-md-writer** — where the per-agent assignment gets recorded.
- **agent-folder-scaffold** — model choice is step 3 of the six-step setup.