Practitioners report Fable 5 planner workflows with Opus, Codex, and HTML logs

• Practitioners are converging on a split-role setup where daniel_mac8's workflow uses Fable 5 as the orchestrator, then hands heavy reasoning to Opus and implementation to smaller coding models.
• skirano's plan-first flow turns Fable output into a markdown spec, then passes that file into Codex with /goal, which matches Anthropic's own guidance around explicit success criteria in Claude Code.
• mattshumer_'s persistent HTML log prompt treats Fable runs like long-lived jobs with a status page, which lines up with Anthropic's Fable 5 prompting guide saying autonomous runs can last for hours and should be checked asynchronously.
• Anthropic is pushing the same basic mental model internally: ClaudeDevs says the Claude Code team now spends less time checking whether work was done and more time checking whether the model picked the right work.
• The catch is that some users still see high cost and awkward routing, as threepointone's cost report questioned a $250 day-long PR and testingcatalog's note flagged prompt classes that transparently reroute to Opus 4.8.

You can read Anthropic's launch post, its Fable-specific prompting guide, and the fallback billing guide. There is also already a GitHub request for a dedicated "fableplan" alias, which tells you how quickly planner-first usage escaped the tweet stream and turned into product feedback.

The strongest pattern in early usage is simple: Fable plans, other models execute. In daniel_mac8's setup, Fable stays on orchestration while Opus handles the reasoning-heavy phases. In skirano's flow, Fable writes an in-depth markdown plan, then Codex gets the file path and /goal for implementation.

That split matches official docs more than the hype cycle does. Anthropic's prompting guide says Fable is unusually strong at long-horizon autonomy, ambiguity, and delegation, while the Claude Code thread from ClaudeDevs' rollout guide explicitly points users to /goal and multi-agent orchestration.

The product-feedback version of the same idea is already live. A June 10 GitHub feature request asks for a fableplan alias so Claude Code can use Fable for planning and Sonnet for implementation without manual model switching.

Long runs are producing their own UI pattern: give the model somewhere to publish progress. mattshumer_'s prompt asks Fable to spin up a persistent HTML page and append timestamped updates with screenshots, turning an autonomous session into a live ops log instead of a blank terminal.

Anthropic's docs explain why that pattern appeared so quickly. The prompting guide says hard requests can run for many minutes, autonomous runs can extend for hours, and teams should add progress indicators and asynchronous checks rather than block on a single response.

Nick Dobos's loop inventory shows what people are starting to wrap around that capability.

• Planning loops: morning briefs, daily planning, weekly review, weekend planning, monthly check-ins NickADobos' planners list
• Maintenance loops: code sweeps, promptbase cleanup, memory and docs consolidation NickADobos' weekly sweeps
• Watcher loops: Sentry crash fixing, analytics-triggered fixes, bug triage from support queues NickADobos' watchers list

Anthropic's best public description of the workflow shift came from its own team. ClaudeDevs says Claude Code users inside Anthropic moved from verifying that the model did the work right to verifying that it was doing the right work.

The rollout thread fills in the mechanics:

• Fable is enabled in Claude Code with /model claude-fable-5 and in the API via the claude-fable-5 model ID ClaudeDevs' model-switch instructions
• Thinking is always on, responses take longer, and Anthropic recommends high effort by default ClaudeDevs on always-on thinking
• Older prompts and skills may be too prescriptive for Fable, so simpler prompting can work better ClaudeDevs on removing old instructions
• Success criteria still matter, exposed as /goal in Claude Code or Outcomes in Claude Managed Agents ClaudeDevs on feedback loops
• Managed Agents can use Fable to delegate work to dedicated agents on smaller models ClaudeDevs on multi-agent orchestration

That list also explains why the planner role keeps showing up in practitioner reports. Anthropic's own Fable 5 guide says the model is more dependable at dispatching parallel subagents and sustaining communication with long-running agents, which is a better fit for supervision than for one-shot inline coding.

The early downside reports are about economics and routing, not raw capability. threepointone's post described spending about $250 on a roughly 10 KLOC PR with lots of testing and intervention, then said smaller steps on earlier models still felt like better value.

Anthropic's official materials confirm that some Fable turns are not really Fable turns. The launch post says prompts in sensitive areas can be answered by Opus 4.8 instead, with safeguards triggering in less than 5 percent of sessions on average, while ClaudeDevs' routing note adds that cybersecurity and biosecurity requests may auto-reroute in the UI and be billed at Opus prices.

The implementation details are already documented in Anthropic's fallback billing guide. Server-side fallback can automatically retry blocked Fable requests on Opus 4.8, annotate the switch with type: "fallback", and expose iteration-level usage so developers can see when the model changed under the hood.

That makes the current practitioner pattern feel less like a quirky prompt hack and more like a hedge against the product's own constraints. Users are already treating Fable as the manager in the loop, then offloading coding, review, or fallback-sensitive work to whatever model is cheaper, steadier, or actually available for the step.