Sonnet 5, Fable 5, and Claude Skills: Choosing Depth Over Noise

By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan

Sonnet 5 as the everyday cheaper model within a point or two of heavier Opus-class siblings, Fable 5 returning after an eighteen-day suspension with more safeguards, and Claude Skills turning Anthropic's prompting guide into something I can invoke with a slash — that combination changed how I choose depth. I stopped chasing the single smartest model for every click and started matching depth to blast radius, cost, and whether I actually needed a one-prompt GTA clone today.

Depth on purpose

Community builds — Pokémon, Monopoly, Minecraft, melody visualisers in one prompt — are wonderful labs. They are a bad default for client Monday. I keep a lab project for one-prompt wonders and a work project where Sonnet-class everyday models draft and Opus-or-Fable-class depth arrives only when evals demand it.

Pricing notes around ten dollars per million input and fifty output versus pricier rivals matter when you multiply by a team. Depth without a budget is a personality tax.

Skills from the official guide

Turning Anthropic's prompting guide into a skill, built on a mid-tier model to save tokens, then invoking /goal so Claude Code plans end-to-end — that is the workflow I teach. Skills encode depth. Chat amnesia discards it.

I version skills. When Fable returns with new safeguards, I re-test the skill pack. Safeguards that change refusals should not surprise a production harness mid-sprint.

Arena honesty

Tests against GPT-5.5, Gemini, Opus — learning roadmaps, one-prompt games, live website redesigns where Fable critiqued before rebuilding — stay in a spreadsheet. Kimi sometimes winning a Fall Guys clone keeps me humble. Not best at everything is a feature of the world, not an insult to my favourite lab.

Ten-test comparisons across GPT, Claude, Gemini, Grok, Kimi are worth repeating quarterly. Loyalty is for people. Models get renewals based on tables.

Sonnet for the week, frontier for the wound

Sonnet 5 within one or two points of Opus 4.8 on many tasks is the everyday driver. I escalate when the cost of being slightly wrong exceeds the token savings — security reviews, tricky refactors, ambiguous product language.

People who run frontier on email rewrites are funding noise. People who run tiny models on auth redesigns are funding incidents.

Agentic coding claims

Describe software, get the finished thing; Fable topping SWE-bench in the notes — treat as a prior, verify on your suite. My suite includes flaky network fixtures and ugly legacy names. Leaderboard suites often do not.

/goal plans are mandatory before long runs. Plans that cannot be pasted into a PR do not start.

Suspension and safeguard awareness

Eighteen-day suspensions and added safeguards are operational events. I keep a fallback model pin and a status check in the morning brief. Building a business process that can only think on one model SKU is how outages become cultural meltdowns.

Safeguards that block a unsafe ask are not enemies. I log them and redesign the ask if it was legitimate work.

What I sell

Skill packs from prompting guides, model routing by blast radius, and quarterly arenas. I do not sell eternal Fable-only shops. Depth is a dial. Noise is what happens when the dial is stuck.

Choice card

Sonnet for volume. Frontier for wounds. Skills for memory. Arenas for humility. Labs for one-prompt joy. Work for shipped truth. That is choosing depth over noise without becoming a bore.

Fallback pin

AGENTS.md lists a fallback model for when Fable or Opus SKUs blip. Sonnet everyday work continues. Skills re-tested after any safeguard change. Depth over noise includes planning for silence from your favourite model.

Lab vs work project split

One-prompt GTA clones live in a lab repo with a play budget. Client work lives elsewhere with Sonnet defaults and frontier escalations. Skills do not cross without review. Depth over noise starts with folder hygiene.

Skill re-test after safeguards

When Fable or Opus returns from a suspension or safeguard bump, I re-run the skill pack on the golden set before client work. Sonnet keeps the lights on meanwhile. Choosing depth includes planning for the deep model to blink.

Everyday driver card

Printed card above the desk: Sonnet for volume, frontier for wounds, skills for memory, arenas quarterly. When noise rises, I read the card before I open a new model tab. Fable demos wait their turn.