tale-mode
A discipline plugin for Claude Code + OpenAI Codex — plan first, verify against the real code, receipts on every decision, independent adversarial review, and a self-armed loop that keeps going until a real check passes. Built by Fable to make any Claude work like a careful senior engineer.
A Claude Code plug-in that scales how much process a coding task gets, from a quick fix to a full review loop for high-stakes changes.
Tale Mode is a plug-in behavior for Claude Code, the AI coding assistant, that tries to make the model work more carefully on complex tasks. It was created after Anthropic discontinued a Claude variant called Fable, and it aims to bring the disciplined working style that Fable was known for to the current Opus model.
The core idea is a throttle that decides how much process a task deserves before starting it. A typo fix gets no ceremony: the model just fixes it. A multi-file refactor triggers a lighter process with verification steps. Anything touching payments, authentication, or security, or anything that would be hard to reverse, triggers the full loop, which includes a separate independent review pass before any code ships.
The full loop runs eight steps: mapping the work and defining what "done" looks like before acting; attaching a cited source or an explicit judgment label to every decision; splitting independent work to sub-agents when appropriate; verifying both internal consistency and against the actual files on disk rather than relying on the model's memory of them; running a self-critique and, for high-stakes tasks, a fresh adversarial reviewer that reads the code independently without the original context; surfacing genuine forks in the work to ask the user rather than guessing; saving progress to a durable file; and naming anything that was not tested or verified.
The skill also ships two slash commands. "/plan-phase" produces a plan for a larger feature and breaks it into independently shippable phases. "/kickoff-phase" picks up one phase in a fresh session, re-verifies the plan against the current state of the code, asks questions at real decision points, and waits for approval before writing anything. The README recommends clearing context between phases because a long session re-reads its entire history on every turn, which slows it down.
Installation is a git clone followed by running an install script. The independent reviewer feature requires a host that supports sub-agents, which Claude Code provides. On the claude.ai web interface, it falls back to a fresh-frame self-review instead.
Where it fits
- Let a coding agent decide automatically how much verification a task needs before starting.
- Trigger an independent adversarial review before shipping payments, auth, or security changes.
- Plan a large feature into independently shippable phases with /plan-phase.
- Resume a phase in a fresh session with /kickoff-phase and re-verify the plan against current code.