Skip to complete skill
Free community skill@theo

Model-strength task router

Independently compiled by XSkills from public material by Theo.

Apply Theo's source-era model rankings and routing defaults, use the named Codex execution paths, judge the result, and escalate when the first route misses the bar.

Skill brief

Is this the right skill?

Use it when

  • You have several available models or agent harnesses with different strengths and costs.
  • You need different routing for mechanical work, user-facing work, computer use, and review.
  • You can evaluate the result and escalate instead of trusting the initial route.

What you get

  • Source-era route and any current substitution
  • Self-contained Codex handoff
  • Observed result and bar decision

Workflow

6 moves
  1. 01

    Use the three source axes

  2. 02

    Apply the explicit defaults

  3. 03

    Route by the dominant requirement

  4. 04

    Use the source's Codex mechanics

  5. 05

    Verify and escalate

  6. 06

    Handle unavailable or cropped mechanics

Complete instruction set

Complete SKILL.md

Read every instruction below or copy the complete file directly into your agent.

Model-strength task router

Source workflow

Use the source's three axes

Rank available routes by cost, intelligence, and taste. Intelligence means how hard a problem the model can take unsupervised. Taste covers UI/UX, code quality, API design, and copy. Theo's reported table is:

ModelCostIntelligenceTaste
gpt-5.5985
sonnet-5557
opus-4.8478
fable-5299

These are Theo's defaults and reflected what he actually paid, not list price.

Apply the routing defaults

  • For bulk or mechanical work, clear-spec implementation, data analysis, and migrations, use gpt-5.5.
  • Anything user-facing, including UI, copy, and API design, needs taste of at least 7.
  • Review plans and implementations with fable-5 or opus-4.8, optionally adding gpt-5.5 as an independent perspective.
  • Never use Haiku under this source policy.

When axes conflict for work that ships, use intelligence before taste before cost. Cost is only the tie-breaker.

Judge and escalate the output

Treat the rankings as defaults, not limits. If a cheaper model's output misses the bar, rerun or redo it with a smarter model without asking. Judge the output, not the price tag.

Use the source's Codex mechanics

Reach gpt-5.5 through the Codex CLI. Use `codex exec` for implementation and `codex review` for reviews, with Theo's `codex-implementation`, `codex-review`, and `codex-computer-use` skills where available. For uncovered investigation or data analysis, run `codex exec -s read-only` with a self-contained prompt.

Claude models run through the Agent or Workflow model parameter. The source begins a thin low-effort Sonnet wrapper that writes a self-contained Codex prompt, invokes `codex exec` through Bash, and returns the result, but the visible instruction is truncated. Do not invent the missing wrapper steps.

XSkills operational adaptations

Preflight the named models, Codex CLI commands, named skills, Agent/Workflow model selection, and source-era defaults. When a model or command is unavailable, ask the operator for the current equivalent and record the substitution. When the thin-wrapper instruction is required, obtain its missing text or use a direct, operator-approved Codex invocation; do not claim the wrapper was reproduced from the cropped source.

XSkills safety adaptations

Routing does not expand authority. Preserve the user's existing permission boundaries for writes, network access, private data, credentials, sends, publishing, spending, destructive work, and computer use. Require explicit approval when the chosen route would cross one of those boundaries.

Return

Return the source-era route selected, any current substitution, the exact self-contained handoff, observed result, whether it cleared the bar, and any escalation.

Source attribution

Independently compiled by XSkills from Theo's post and attached CLAUDE.md policy for choosing models in workflows and subagents. Theo did not author or review this skill. Capability and substitution handling, cropped-wrapper fallback, and explicit permission boundaries are XSkills adaptations. Attribution does not imply endorsement.

Keep exploring

Related agent skills

Browse ai workflows skills

Turn an X workflow into your own skill.

Paste a public X post and get a free SKILL.md.

Build one free

Evidence and package

Source, changes, and version

Source

XSkills independently compiled this from public material by Theo. Attribution does not imply endorsement.

Open source evidence ↗

What XSkills changed

XSkills additions are labeled in the complete SKILL.md.

Limitations

  • Do not hard-code source-era model names, rankings, prices, or availability as current facts.
  • Do not pass secrets or sensitive data into a delegated model without an approved handling path.
  • Do not grant write, shell, network, or computer permissions when a read-only route can complete the task.

Version

Package version
Not assigned
Fingerprint
9dcfcde0a64b
Last reviewed
Not recorded
License
Not specified