Personal operating guide · Research snapshot: 03 Aug 2026

Pick the right AI setup before you spend the work.

A practical routing system for Codex, Claude Code, OpenCode, and OpenRouter—built around the work you actually do: websites, serious workbooks, and software.

Route my task  ↓Build a task brief

The rule is not “use the strongest model.” It is: use the cheapest setup that can safely reach a verifiable result.

1. Which app to open2. Model strength3. Acceptance proof
01

Classify the work

Choose the actual deliverable and stage, not a model from habit.

02

Choose where to work

Pick the app to open first. This is separate from choosing the model inside it.

03

Choose a tier

Utility, builder, or architect—based on the cost of a wrong answer.

04

Demand proof

A live page, workbook checks, or tests make “finished” measurable.

Decision tree

Start here, every time.

Answer four questions. The recommendation identifies where to begin, what capability class to choose, and what must be checked before you call it complete.

Describe the task

No account details needed
How far does this change reach?This means how many parts of your project may be affected. There is no wrong answer—if you are unsure, choose the last option.
Choose High if an error could affect money, private information, security, or something that would be difficult to undo.

Capability tiers

The tier comes after the task.

Use labels in the model picker you actually have. Names, limits, and prices move; this decision logic should not.

Tier 1 · Utility

Scout

Fast, cheap help for contained work where you can quickly verify the answer.

Examples: Codex Luna where available, Claude Haiku, or a low-cost OpenRouter model that has passed a small trial.

  • Find references, explain one function, rename a field
  • Generate test cases or formula examples to review
  • Clean data, copyedit docs, summarize bounded logs
Tier 2 · Default

Builder

Your daily executor for a clear specification, multi-file implementation, and ordinary debugging.

Examples: Codex Terra or the current Claude Sonnet family—whichever your subscription exposes.

  • Build a responsive landing page from a written brief
  • Create an Excel model with checks, formulas, and charts
  • Implement a planned software feature and its tests
Tier 3 · Architect

Reasoner

Reserve for ambiguity, real tradeoffs, unfamiliar systems, or a mistake that would be expensive to repair.

Examples: GPT-5.6 Sol in Codex or Claude Fable 5 in Claude Code. Fable 5 is Anthropic’s most capable widely released model; use it for the hardest reasoning and long-running agent work.

  • Make an architecture, data, or migration plan
  • Untangle state, concurrency, or non-obvious defects
  • Perform a security-sensitive design and review pass

Architect → Builder → Verifier

The safest handoff is an artifact, not a chat recap.

For sizeable projects, make the architect write a compact plan to disk: goal, constraints, affected files, success checks, and open questions. The builder executes that plan in small verified slices.

This preserves useful work if you switch apps, run out of allowance, or come back tomorrow. It also keeps expensive reasoning away from routine edits.

Generate a handoff-ready brief
1

Frame

State the deliverable, non-negotiables, affected area, and acceptance test. If those are unclear, start with an architect.

2

Plan

Ask for options, risks, and a file-by-file or sheet-by-sheet plan. Save the decision, not pages of internal reasoning.

3

Build

Give the builder the selected plan. Request small changes, visible outputs, and a test after each meaningful milestone.

4

Verify

Use an independent check when impact is high: browser inspection, formula audit, tests, or a second-model review with a rubric.

Your work, translated

Three playbooks you can reuse.

These are operating patterns, not rules requiring a particular vendor. Pick the current model family available in the tool you are using.

Task brief builder

Give any model a good starting point.

Fill this in before you begin. It creates a concise handoff prompt containing what a capable agent needs without forcing it to rediscover the assignment.

Include exact files, examples, and known errors when you have them. Do not paste credentials or private data.

Copy into the new task

Updated now

Efficiency guardrails

Spend attention where it changes the answer.

Context and agent loops drive cost. The practical answer is better framing and checkpoints—not blindly chasing the lowest price.

Before starting

Write the acceptance test first.

“Make the dashboard better” causes exploration. “Make totals match the source table, preserve filters, and show the variance chart” gives the agent a checkable target.

While building

Keep one feature’s context together.

Do not reset a useful thread just because it is long. Save a compact handoff at a natural boundary, then begin fresh only when prior context is more confusing than helpful.

For OpenRouter

Test before you standardize.

Evaluate a candidate on one representative task with a known check. Record quality, time, tool behavior, and actual cost. A cheap model needing three rescues is not cheap.

For subscriptions

Use included work first.

Subscriptions are excellent for normal iterative development. Keep gateway credit for purposeful experiments, a specific unavailable capability, or overflow—not every routine task.

For large context

Stabilize reusable context.

Put durable instructions and known facts in a concise project guide. Keep dynamic status, timestamps, and one-off requests separate so compatible providers can reuse cached prefixes.

When stuck

Stop, preserve evidence, then escalate.

Do not keep saying “try again.” Save the error, diff, failing test, and attempted fix. Give that packet to a stronger model or a clean task with a precise diagnostic question.

Keep it current

Principles here; changing facts at the source.

Model names, access, rates, and limits move fast. This site avoids pretending that one hard-coded price table will stay correct. Review these official references when changing subscriptions, selecting a new model, or configuring a gateway.