asets-gobizit/token

token — Model picker for Claude Code

Show the model comparison table, present a picklist, and persist the user's choice to ~/.claude/settings.json.

When to use this skill

• User types /token
• User says "token" alone as a command
• User asks "which model should I use?", "compare models", "change model", "pick model", "switch model"
• User wants to see the cost-vs-capability tradeoff between available models

Instructions (run these in order)

Step 0 — Auto-detect current models (ALWAYS DO THIS FIRST)

DO NOT use hardcoded model names. Instead:

1. Primary source (live docs): Use WebFetch to check https://platform.claude.com/docs/en/about-claude/models/overview — extract the latest model names, version numbers, and model IDs from the "Latest models comparison" table. This is the authoritative source and may be ahead of the system prompt.
2. Fallback (system prompt): If WebFetch fails, read your own system prompt — it contains a line like: "The most recent Claude model family is Claude X.Y. Model IDs — Opus X.Y: 'claude-opus-x-y', Sonnet X.Y: 'claude-sonnet-x-y', Haiku X.Y: 'claude-haiku-x-y'"
3. Also check: "You are powered by the model named ..." — this tells you which model the current session is using.
4. Use these extracted values to populate the table, picklist, and settings mapping below.
5. If the live docs show a newer model than the system prompt lists, use the docs version — the system prompt may lag behind new releases.

IMPORTANT: Always show the version number in display names. Write "Opus 4.7", "Sonnet 4.6", "Haiku 4.5" — NEVER just "Opus", "Sonnet", "Haiku" without the number. The user wants to see exactly which version they're picking.

This ensures the skill automatically reflects any model upgrade without needing a manual edit — even when the system prompt hasn't been updated yet.

Model tiers (these roles are stable even as versions change):

| Tier | Example display name | settings.json alias | Typical ID pattern | |---|---|---|---| | Opus (1M context) | Opus 4.6 (1M context) | Use full model ID with [1m] suffix | claude-opus-{version}[1m] | | Sonnet | Sonnet 4.6 | sonnet | claude-sonnet-{version} | | Haiku | Haiku 4.5 | haiku | claude-haiku-{version} | | Sonnet (1M context) | Sonnet 4.6 (1M context) | Use full model ID with [1m] suffix | claude-sonnet-{version}[1m] |

Step 1 — Show the comparison table

Build and display this table using the auto-detected model names from Step 0:

---

Claude Code model comparison

| # | Option | What it is | Best for | Watch out | |---|---|---|---|---| | 1 | {Haiku + version, e.g. "Haiku 4.5"} | Smallest, fastest, cheapest | Quick lookups, subagent work, "just answer this" | Not smart enough for real work. Frustrating on anything multi-step. | | 2 | {Sonnet + version, e.g. "Sonnet 4.6"} ⭐ | ~95% as smart as Opus, way more efficient | Daily driver — coding, refactors, normal Q&A, file edits, almost everything | 200K context is plenty for 95% of sessions. Only feels limiting on huge transcripts/codebases. | | 3 | {Opus + version, e.g. "Opus 4.6"} (1M context) | Most capable model, biggest brain | Hard architecture decisions, deep debugging, strategy, anything multi-step where mistakes cost time | Burns through Max plan usage ~5x faster than Sonnet. You'll hit weekly limits sooner. | | 4 | {Sonnet + version, e.g. "Sonnet 4.6"} (1M context) | Same Sonnet but 1M context, EXTRA COST | One specific case: dumping a giant codebase or transcript into one session | Extra-billed — outside your plan. Real $$$. |

After the table, show: Currently running: **{model from system prompt}**

---

Step 2 — Present the picklist

Use the AskUserQuestion tool with options built from the auto-detected model names:

• header: "Pick model"
• question: "Which model do you want to switch to? This will be saved to settings.json for the next session."
• multiSelect: false
• options: (use auto-detected names for labels)
  1. label: {Haiku + version, e.g. "Haiku 4.5"} — description: Fastest. Skip for real work.
  2. label: {Sonnet + version, e.g. "Sonnet 4.6"} — description: RECOMMENDED daily driver. Best efficiency.
  3. label: {Opus + version, e.g. "Opus 4.6"} (1M context) — description: Most capable. Burns Max-plan usage fast.
  4. label: {Sonnet + version, e.g. "Sonnet 4.6"} (1M context) — description: EXTRA COST — billed outside your plan! Big inputs only.

Step 3 — Map the user's choice to a model string

Use the auto-detected model IDs from Step 0:

| User picked | Write to settings.json "model" | |---|---| | Opus (1M) | {opus-model-id}[1m] (e.g., claude-opus-4-6[1m]) | | Sonnet | sonnet (always use alias — it auto-resolves to latest) | | Sonnet (1M) | {sonnet-model-id}[1m] (e.g., claude-sonnet-4-6[1m]) | | Haiku | haiku (always use alias — it auto-resolves to latest) | | Cancel / Other dismiss | (do nothing, exit) |

Step 4 — Update settings.json

1. Read C:/Users/dansk/.claude/settings.json
2. Use the Edit tool to update the "model": "..." line with the new value
3. DO NOT touch any other field (permissions, additionalDirectories, statusLine, hooks, skipDangerousModePermissionPrompt)
4. If the "model" field doesn't exist yet, add it as the last field of the root object before the closing brace

Step 5 — Confirm and instruct

Print this exact confirmation (substituting <MODEL_LABEL> with the option the user picked):

Saved — next Claude Code session will use <MODEL_LABEL>.

To switch the current session immediately:

• Type /model and pick the same option, OR
• Reload VS Code: Ctrl+Shift+P → Developer: Reload Window

If the user picked Cancel, just say:

Cancelled — no changes made.

Constraints

• Never edit any file other than ~/.claude/settings.json
• Never modify other fields in settings.json
• Never assume what the user wants — always show the table and picklist first
• Never hardcode model version numbers — always read them from the system context
• If AskUserQuestion fails or the user types a free-form answer instead of picking, fall back to interpreting their text and confirming with one yes/no question before editing