fable5-skills

fable5 - Loop Engineering with Claude Fable 5 and Mythos 5

Working examples of loop engineering with Anthropic's Claude Fable 5 and Mythos 5: a deep research agent, an agentic SOC triage loop, and a Mythos scope gate for approved defensive security work. The shared loop anatomy is trigger, rules, executor, fresh-context verifier, memory, and stop rules.

Companion repo to my Substack series:

• Part 1 — Claude Fable 5: What Changed, and How to Stop Prompting It Like Opus
• Part 2 — Loop Engineering, or How I Stopped Writing Prompts (paid)
• Part 3 — Memory Engineering (paid)
• Part 4 — Ten Weekend Security Projects, or Plan Expensive, Build Cheap (paid)
• Part 5 — Auditing a Third-Party Fable 5 Tool, Then Building Our Own (paid)
• Part 6 — Ten Fable 5-Native Agent Skills (paid)
• Part 7 — Moving a Production Fleet Off Opus 4.8 (paid)
• Part 8 — The Refusal Your Error Handler Can't See (paid)
• Part 9 — Don't Pay Twice for a Fallback (paid)
• Part 10 — The Effort Dial Is a Cost Dial (paid)

What's here

• deep_research/ — a bounded research loop: decompose the question at xhigh effort, search each angle with the server-side web search tool, verify every claim with three fresh-context adversarial voters, synthesize only what survives.
• agentic_soc/ — an alert triage loop: enrich and classify each alert, re-score it with a fresh-context verifier that sees only the artifacts, escalate or close, and distill lessons from verifier-confirmed verdicts only. Read-only by design; containment stays behind a human gate.
• memory/ — memory engineering: a client-side handler for the official memory tool (memory_20250818) with traversal guards and size caps, and a startup scan that quarantines lessons that command instead of describe (the anti-poisoning gate).
• templates/ — the loop instruction blocks and the memory scaffold, as paste-ready markdown.
• job_packet/ — the shared library behind the ten weekend projects below: prep (cheap model), plan_or_audit (Fable 5, returns a fallback category when classifiers reroute the request), execute (cheap model), verify (fresh-context grader).
• projects/ — ten weekend projects for security architects, agentic AI engineers, and CISOs, each following prep → plan/audit → execute → verify. Every project ships with generic sample data, so none of them need your actual codebase to run.
• fable_scanner/ — a clean-room Fable 5 hygiene tool with three subcommands: scan (lint CLAUDE.md/skills/agent configs for the anti-patterns in Anthropic's own Fable 5 prompting guide), canary (regression canary — a fixed prompt set run periodically, alerting on refusal-rate or latency drift), and monitor (refusal-category and cost analytics from local Messages API response logs, no network calls). Independently designed and written; see Part 5 for why and how it was audited before anything from the third-party tool that inspired it was trusted.
• skills/ — ten Fable 5-native Agent Skills (SKILL.md files, installable as a Claude Code plugin): effort routing, scope fencing, no gold-plating, evidence-bound progress, act-on-sufficiency, verifier subagents, autonomy contracts, lesson memory with an anti-poisoning gate, reader-first summaries, and refusal-aware API calling. Independently designed after auditing a third-party skills collection for Part 6.
• migration/ — a linter that diffs an existing Opus 4.8 request body against Fable 5's requirements: it flags the sampling params (temperature, top_p, top_k) and manual thinking config that now return a 400, carried thinking blocks that bill as dead input tokens after a model switch, and a missing fallbacks parameter. Returns (field, severity, message) tuples so CI can fail a build on any error. Part 7.
• refusal/ — the wire-level refusal handler: send with the server-side fallback wired, then a read function that branches on stop_reason (never on content, which is empty on a refusal) and reads usage.iterations to report whether Fable answered directly or a fallback caught a decline. Part 8.
• fallback_credit/ — manual fallback-credit redemption for custom retry loops that lack the SDK middleware: redeem the one-time fallback_credit_token on an exact-body-match retry so a hand-rolled fallback isn't billed for two cache writes, plus a served_by_sticky check that separates a sticky-routed turn from a real fallback. Part 9.
• effort/ — a router that maps each task's kind to one of the five effort