Always-on operational guardrails, model-independent. Apply on EVERY task and EVERY model whether or not fable-mode's staged loop is running. Three rules: verify-before-flag, warning batching, and find-and-replace safety. Trigger whenever a response would flag a problem, raise a warning, or edit files with search-and-replace.
インストールする前に、エージェントに与えられる指示の中身を確認できます。
Family version: v6 (2026-10-10).
Three rules extracted from fable-mode's operational section so they apply everywhere, not only inside the staged loop. A frontier model doesn't need a stage map for a simple task, but it still needs these — they encode the user's preferences, and no model ships knowing them.
Before flagging any problem — verify it actually exists. Grep, diff, run it, or check the source directly. Never report a problem that hasn't been confirmed present.
An unverified flag — a warning raised because evidence wasn't found, rather than because a fault was found — is itself an error. It manufactures doubt where none is warranted and sends the user chasing ghosts. Absence of evidence is not the finding. Confirm, then flag.
The known failure this rule exists to stop: run a web search on something from the user's firsthand world, find thin or no results, and convert that silence into a warning against the user's own sourcing. Web silence is never grounds for a warning. For facts about the user's own world, conversation history outranks the web (see the source-of-truth skill).
A capability flag follows the same standard. "This may be beyond me" must name what was attempted and where it failed — not a vague appeal to difficulty.
Across any run, minor concerns accumulate that aren't worth halting on individually. Keep a running count. At the threshold — default three, tunable if the user sets a different number — stop and surface all of them to the user at once before continuing.
Rationale: three small things pointing the same direction usually mean one real problem worth a decision. Below threshold, keep working; a drip of trivial caveats is noise. At threshold, batch them — one interruption with full context beats three fragmentary ones.
A concern that independently meets the verify-before-flag bar and is material on its own does not wait for the threshold. The threshold governs minor concerns only.
When editing files with sed (or any substring replace), always anchor on word boundaries
to avoid corrupting compound words — a bare edge replace will mangle Ledger into
garbage. Use \bword\b, not bare word.
After any sed pass on a file, grep for glued or malformed compound words before presenting the result. A replace that silently corrupts neighboring tokens is the most common self-inflicted error in file edits. The check is cheap; the silent corruption is not.
Preferred order of tools: targeted string-replace on a unique anchor > word-boundary sed > bare sed (never). If the string to replace isn't unique in the file, widen the anchor until it is — do not replace-all and hope.
fable-mode's staged loop is optional and gated on task size. These guardrails are not optional and not gated. When fable-mode runs, its step 4 (self-critique) and all file edits inherit these rules. When fable-mode doesn't run, these rules apply anyway. The per-model runners — fable-opus and fable-sonnet — inline these rules verbatim into every subagent briefing: cloud subagents could load this skill (2026-10-10), other surfaces are unchecked, and the inline copy is shorter than a skill load.
まだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Automatic final verification pass on completed deliverables. Runs at the end of every fable-mode run — those skills invoke it before delivery. Also triggers when the user says "double check this". Uses a panel of fresh checker agents when spawning is authorized (otherwise a cold self-check) to verify the finished output — not to redo the task. Any FAIL gets fixed and re-checked before delivery.
日本語の概要は準備中です。原文の説明を表示しています。
Run fable-mode execution discipline on Claude Fable 5.1 — the top of the escalation ladder and the strongest staged run available. Routes the task to the @fable-orchestrator agent with its model overridden to Fable (`model: "fable"`), which stages the work, delegates artifact production to @fable-worker-sonnet / @fable-worker-haiku, and cold-checks deliverables with @fable-verifier. Trigger when the user explicitly asks for thorough/systematic/"deep work" handling on the strongest model ("fable on fable", "stage this on fable", "deep work mode, fable", "run this on 5.1"), or when a fable-opus run reports the task exceeds Opus's ceiling. Benchmarked 2026-09-02: zero correctness delta on Fable 5.1 at ~1.1× tokens, ~2× wall — a provenance runner, not an accuracy runner. Do NOT use for ordinary single-pass tasks — prefer fable-opus, fable-sonnet, or fable-haiku when the task doesn't need peak reasoning. Fable is plan-included; bulk work still routes down-tier.
日本語の概要は準備中です。原文の説明を表示しています。
Run fable-mode execution discipline on Claude Haiku. Routes the task to the @fable-worker-haiku agent, whose definition carries the staged loop with tightened verification (no bare "unverified" allowed) and an escalate-don't-improvise rule. Trigger when the user explicitly asks for thorough/systematic handling run cheaply or fast ("fable on haiku", "deep work mode but cheap", "stage this on haiku"). For bulk mechanical work. Do NOT use for tasks needing synthesis — benchmark note: at n=1 the skill's effect on Haiku swung both directions (+25 / −17); route quality-critical work to fable-sonnet instead.
日本語の概要は準備中です。原文の説明を表示しています。
Enforces staged execution discipline on large tasks: a written stage plan, parallel delegation where the runtime supports it, a failable verification check at each stage, and a skeptical self-review before delivery. Trigger when the user explicitly asks OR when the task objectively spans multiple files, sources, or sessions. Do NOT trigger on ordinary multi-step requests that a direct attempt handles fine (autonomous triggering only; explicit invocation always runs).
日本語の概要は準備中です。原文の説明を表示しています。
Run fable-mode execution discipline on Claude Opus — the top of the escalation ladder. Trigger when the user explicitly asks for thorough/systematic/"deep work" handling on Opus ("fable on opus", "stage this on opus", "deep work mode, opus"). When a task exceeds Opus's ceiling, report that to the user rather than escalating further.
日本語の概要は準備中です。原文の説明を表示しています。
Run fable-mode execution discipline on Claude Sonnet. When invoked by slash command or by name, spawns a Sonnet subagent that follows the staged loop — stage plan, parallel delegation, a failable verification check at each stage, and a skeptical self-review before delivery; a wording-only request runs the loop inline on the current model. Trigger when the user explicitly asks for thorough/systematic/"deep work" handling on Sonnet.
日本語の概要は準備中です。原文の説明を表示しています。