thedixitjain Fable 5 Orchestration
Fable = architect only (~5–15% of the work: high-level structure + ambiguous judgment). Opus 4.8 does everything else — research, execution, verification. No Sonnet.
The loop: Opus 4.8 researches → Fable architects → Opus 4.8 executes → Opus 4.8 verifies. Fable only touches the middle step; you pay its rate only for the thinking that needs it.
Without a Claude Code subscription you can drop Fable entirely and run GPT-5.6 Sol as the architect instead — it's close enough now that the cost math often wins. See Add GPT-5.6 Sol (Codex) below.
Wire it (pick one)
- Advisor mode — set active model to Opus 4.8, then
/advisor fable. Opus writes everything; Fable steers when Opus is stuck. One clean session. Simplest to run/film. - Architect + delegate — Fable writes the plan, then parallel Opus 4.8 subagents (or a separate Opus 4.8 session) execute it. Bigger spectacle: a rail of Opus agents building while Fable conducts.
Effort = the cost lever
/effort low|medium|high|xhigh|max. Cap at high — avoid xhigh and max. On Fable they usually make things worse: it over-reasons, runs in loops, and ships overdone code with too many changes at a much higher cost. Higher effort doesn't let it work longer or solve harder problems — it just thinks more per step. Use low/medium for routine work (Fable low often beats prior models at their max, for a fraction of the cost); high is the ceiling.
Add GPT-5.6 Sol (Codex) — solo, or in tandem with Fable
Fable is still the sharpest orchestrator, but GPT-5.6 Sol is now right behind it for a fraction of the cost (roughly $5/$30 per million tokens vs Fable's $10/$50). Two ways to run it, depending on whether you already pay for Claude Code.
No Claude Code subscription — run Sol on its own. Use the Codex CLI directly; Sol plans and builds the whole thing itself:
codex exec -m gpt-5.6-sol "<the build task>"
gpt-5.6-sol is the model id. Keep effort matched to difficulty (medium/high for most work; skip the top ultra tier, which burns tokens for little gain).
Have Claude Code + a ChatGPT account — run them in tandem. Sign Codex into Claude Code with OpenAI's official plugin, then let Fable orchestrate and delegate implementation to Sol. Needs a ChatGPT plan (or OpenAI API key) and Node 18.18+:
/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins
/codex:setup
Then sign in with !codex login (ChatGPT account or API key); usage counts against your Codex limits. Now Fable architects up top and delegates implementation to Opus 4.8 agents and to Sol via the plugin's /codex:* commands.
Which you pick is a budget call: on cost alone Sol-only is the easy default; if you already have Claude Code, Fable as the top conductor with Sol underneath is still a touch sharper all around.
Paste-in prompt kit
Drop into the Fable setup's system prompt / CLAUDE.md / skills. Short instructions steer better than rule-piles — state each behavior once.
Act, don't overplan:
When you have enough information to act, act. Don't re-derive established facts, re-litigate a settled decision, or narrate options you won't pursue. If weighing a choice, give a recommendation, not a survey.
No unrequested tidying/refactor:
Don't add features, refactor, or introduce abstractions beyond what the task requires. Do the simplest thing that works. Only validate at system boundaries; don't handle scenarios that can't happen.
Delegate to subagents:
Delegate independent subtasks to subagents and keep working while they run. Prefer async over blocking until each returns. Intervene only if a subagent goes off track or lacks context.
Ground progress claims:
Before reporting progress, audit each claim against a tool result from this session. Only report work you can point to evidence for; if unverified, say so. If tests fail, say so with the output.
State the boundaries:
When the user is describing a problem or asking a question rather than requesting a change, the deliverable is your assessment — report and stop; don't fix until asked. Before a state-changing command, check the evidence supports that specific action.
Verify with fresh eyes:
Every [interval], verify your work with a fresh subagent against the specification.
Give the why:
I'm working on [larger task] for [who it's for]. They need [what the output enables]. With that in mind: [request].
Autonomous run:
You're operating autonomously; the user can't answer mid-task, so don't ask "Want me to…?". For reversible actions that follow from the request, proceed. Before ending your turn, if your last paragraph is a plan, a question, or a promise ("I'll…"), do that work now. End only when done or blocked on user-only input.
Memory:
Store one lesson per file, one-line summary at top. Record corrections and confirmed approaches and why they mattered. Don't duplicate what the repo records; update rather than duplicate; delete notes that turn out wrong.
Hard don'ts
- Never tell Fable to show/echo/"explain your reasoning" in the response — trips the reasoning-extraction classifier and silently reroutes to Opus 4.8. Audit system prompts/skills for it. For visibility, read the
thinkingblocks instead. - Don't over-instruct. Strip legacy rule-piles; a brief instruction steers Fable better.
- Don't block on subagents — bottlenecks on the slowest, burns context.
- Don't predefine subagent archetypes (reviewer, adversarial, explorer). Fable invents the right roles per task — prescribing them just narrows it.
- Don't leave effort on high for routine work.
- Don't send research/boilerplate to Fable — that's Opus 4.8 work; hand Fable the results.
Pre-run checklist
- Is Fable actually needed for this step, or does Opus 4.8 take it? (~5–15% Fable.)
- Effort matched to difficulty, not defaulted to high?
- Advisor mode or architect+delegate wired?
- System prompt free of any "show/explain your reasoning" language?
- Delegate + verify + boundaries blocks in place?
Source: per-simmons/fable-orchestration → SKILL.md