R068 · Harness tips › Harness tips

Plan on the expensive model, hand off to the cheap one: Hassid's Fable-to-Opus brief

Read the original

harness-configpromptingprompt-cachinginference-costhandoff

Sources: Ruben Hassid's LinkedIn post (2026-11-13) and his newsletter issue Prompt 5.2. (How to AI, 2026-11-13). Claims about caching and model behavior are attributed in the newsletter to Anthropic's docs and help center.

The advice

Use the newest, most expensive model (Claude Fable 5.2) for one big job: planning and architecting the whole "how". Then move execution to the cheaper everyday model (Claude Opus 5.5) through a written brief in a new chat, not by switching models mid-conversation. The newsletter adds that Fable's inclusion in standard Team seats is time-limited, after which it bills against usage credits. That makes the cost split practical, not theoretical.

Why not just switch models in the chat

Two reasons, both sound:

  1. Caches are model-scoped. A long chat stays cheap because its prefix is cached. Switching models means the new model reads the whole history cold, at full input price. ("Rereads from zero" is the newsletter's phrase. Mechanically, it is a cache miss on the whole prefix.)

  2. Reasoning doesn't transfer. The second model sees the visible transcript, not the first model's thinking. So the plan's rationale is lost unless it is written down.

The fix for both is the same: have the planning model write an explicit, self-contained brief, and start fresh.

The handoff brief (structure)

The newsletter's handoff prompt asks the planning model to write for a reader who saw none of the chat, copying exact names, numbers, dates, links and file names, keeping only final versions and adding nothing new. Its sections, paraphrased:

PartPurpose
Goal + "done when"One sentence, then the checks that define success
Next stepThe first action the executor can take without the human
DecidedEach final decision with a short reason, plus rejected options and why
FactsEverything the work depends on
Done hereWhat is finished, so it isn't redone
Ask me firstWhat needs human approval, and why
OpenOnly what is still undecided; no new questions
ChecklistEvery remaining step as tick boxes
Executor rulesRead every relevant file and connected app first; work the checklist; keep going instead of offering to continue; stop only for approvals; don't reopen decisions but flag contradictions; end with a short recap

The executor is then started in a new chat with the brief attached and a one-line instruction to take it from there.

Why it works, in engineering terms

This is context compaction done by hand, with a schema. The brief is a lossy but deliberate summary in which:

The same structure works for handing off between agents, between sessions, or from an agent to a human.

Caveats

Checklist for using the pattern

  1. Does the brief state "done when" in checkable terms?

  2. Are rejected options listed, with reasons?

  3. Are approval gates named, especially for deletes, sends and spend?

  4. Would a newcomer understand every name, path and number without the original chat?

  5. Is the brief short enough that the executor reads all of it?

Bottom line

Practical, cost-aware advice with a reusable artifact: a structured handoff brief. The underlying reasons (model-scoped caches, reasoning that doesn't transfer) are correct. The brief template is worth keeping even if you never split work across models.

Connections

See R068 on the knowledge map

Part of the Research notes on nowucca.com. Figures and claims are the source's own unless the summary says otherwise.