Fable Prompt Craft — Briefing Weaker Models

A weaker model with a frontier-quality briefing beats a frontier model with a lazy one. The gap between model tiers is mostly: judgment under ambiguity, edge-case anticipation, and knowing when to stop. All three can be supplied in the prompt. That's what this skill does.

The core principle: move judgment into the brief

Frontier models fill gaps correctly; weak models fill gaps plausibly. So a brief for a weaker model must contain:

1. The decision already made. Don't ask it to choose an architecture — give it the architecture and let it implement. Save open-ended design for strong models or humans.
2. The gold example. One concrete example of correct output beats three paragraphs of description. Two examples (a normal case + an edge case) beat almost anything.
3. The failure modes, named. "Do NOT put the API key in VITE_* vars. Do NOT call the email provider synchronously inside the save hook." Weak models avoid named traps reliably. They discover unnamed traps almost never.
4. The stop condition. "Done means: X passes, Y exists, Z returns 200." Without it, weak models either under-deliver or gold-plate.

Structure of a strong task prompt (in order)

CONTEXT   — the system, the stack, what exists already (paths!), what must not change
TASK      — one clear objective, then numbered sub-steps in execution order
CONSTRAINTS — locked rules, named failure modes, things NOT to do
EXAMPLE   — gold input/output pair, or a snippet of the existing pattern to imitate
VERIFY    — exact commands/checks that must pass before claiming done
REPORT    — what the final message must contain

Why order matters: models weight the start and end. Context first (frames everything), verification last (the last thing read is the first thing done before replying).

Rules that pay for themselves

• Be self-contained. The session/agent has none of your context. Every path absolute, every term defined, every "the usual way" spelled out. Test: could a competent stranger with no chat history execute it?
• Numbered steps for sequence, bullets for constraints. Mixing them loses ordering information.
• State priority explicitly when goals conflict: "If correctness and speed conflict, choose correctness."
• Quantify vagueness away. "A few" → "3". "Robust" → "handles empty input, duplicate delivery, and 10MB payloads." "Fast" → "under 200ms".
• Give escape hatches. "If X is not found within 5 minutes, note where you looked and move on" — otherwise agents burn an hour on a dead end or, worse, fabricate.
• Forbid the specific lies. Weak models round "mostly works" to "works". Require: "Report failures verbatim. If a step was skipped, say so. Never claim a test passed without running it."
• One task per prompt. Two loosely-related tasks in one brief → the second gets 20% effort. Chain sessions instead.

For agents/automations specifically

• Tell the agent what its output IS: "Your final message is the deliverable — include the full report, not a summary of where you put it" (or the opposite: "write to file X, reply with 5 lines").
• Structured output: specify the exact schema/format and give one filled-in example row.
• Parallel agents must have disjoint write targets. Two agents writing one file = corruption. Assign each its own output path.
• For research agents: require a link, an excerpt, and a summary per source, and cap the scope ("top 10 competitors, not exhaustive").

Session handover briefings (the handover.md protocol)

A handover is a prompt for a future model that knows nothing. It must contain, concretely:

• What is DONE (verified, with how it was verified) vs what is IN PROGRESS (exact state: "endpoint written, not yet wired to the email provider") vs NOT STARTED.
• Next 1-3 steps with exact files, commands, expected outcomes — executable without archaeology.
• Rules introduced this session (things a fresh model would violate).
• Dead ends already explored ("don't try X — fails because Y"). This is the highest-value, most-omitted section: it prevents the next session from burning an hour rediscovering a wall.
• Pivots and why.

Anti-patterns

• Politeness padding and role-play framing ("You are a world-class engineer…") — costs tokens, changes nothing. State the task.
• Prompting a weak model to "think carefully" instead of giving it the checklist a careful thinker would use.
• Ten constraints of equal weight. Rank them or mark the three that are non-negotiable.
• Asking for creativity AND compliance in the same sentence. Split: generate options first, then converge under constraints.
• Re-prompting a failed task with "try again, but better." Diagnose which part of the brief it missed and reinforce THAT.