Skill Fablize for Hermes

English ·

简体中文·

Português (BR)

A port of fablize for the Hermes Agent ecosystem. This project adapts fablize's verified procedures — verification grounding, multi-story evidence gating, investigation protocol, and early-stop prevention — from Claude Code's plugin system to Hermes Agent's skill system.

fablize makes any agent see a task through to the end — with evidence and verification — as procedure, not as luck.

fablize was born from a controlled comparison of Claude Opus and "Fable 5" (19 A/B runs + 26 real sessions, ~1,500 tool calls). The finding: on closed, answer-bearing work (code, logic, builds) the two models were effectively tied. The gap was in depth — following an implication one step further — and that turned out to be model capability, not injectable.

But the procedure of good work — actually running what you build, decomposing multi-step tasks with evidence checkpoints, investigating bugs systematically — does transfer. fablize ships only the procedures whose effect was verified.

The core behavioral ruleset, structured as a Hermes Agent skill (YAML frontmatter + markdown). Covers:

Foundation rules— always-on: lead with outcome, ground claims in tool results, no early stops, confirm destructive actions** Multi-story loop**— decompose 2+ story tasks → complete with evidence → verification gate on the final story** Deep investigation**— reproduce → 3+ competing hypotheses → trace causal chain → verify before/after → report rejected hypotheses** Verification grounding**— run render/executable artifacts, observe, fix, re-run before declaring done** Capability ceiling**— escalate honestly when the model's ceiling is the blocker

A self-contained, stdlib-only Python script for multi-story decomposition with a verification gate. State persists in ~/.hermes/fablize/

and survives session death.

goals.py create --brief "Build a dashboard" \
  --goal "design::Create the HTML layout" \
  --goal "implement::Add chart visualizations" \
  --goal "verify::Run and verify the dashboard renders correctly"
goals.py next
goals.py checkpoint --id G001 --status complete --evidence "Layout created with sidebar and main panel"
goals.py status

— The grounding loop: run, observe, fix, re-run. For HTML, SVG, games, UI, charts, animations.packs/verification-grounding-pack.txt

— The systematic debugging protocol: reproduce, competing hypotheses, causal chain, verify before/after.packs/investigation-protocol.txt

setup.sh [--skill-only|--local|--global]

— Install the skill into Hermes and optionally inject the operating block intoAGENTS.md

orCLAUDE.md

uninstall.sh [--skill-only|--local|--global|--all]

— Reverse the setup

git clone https://github.com/zeus/fablize-for-hermes.git
cd fablize-for-hermes
bash setup/setup.sh --skill-only

Then load the skill in your Hermes session:

skill_view(name='fablize')
bash setup/setup.sh --local    # injects block into ./AGENTS.md
bash setup/setup.sh --global   # injects block into ~/.claude/CLAUDE.md
bash setup/uninstall.sh --all

Trigger: load thefablize

skill, or trigger inline when the agent detects 2+ stories, debugging, or render artifacts2+ stories→ decompose + verification gate viagoals.py

Debugging→ investigation protocol (reproduce → hypotheses → causal chain)** Render artifact→ verification grounding (run → observe → fix → re-run) Hard task→ adaptive thinking + escalate to stronger model if stuck Capability ceiling**→ escalate honestly (the harness does not fake capability)

• It cannot raise model capability. Open-ended creative quality and self-driven discovery are out of reach — that is a model-choice decision, not a harness one.
• The effect numbers come from the original fablize small-family self-measurement. The direction is solid; the decimals are not asserted.
• The early-stop rule relies on agent self-enforcement, not a deterministic hook. It is a guideline, not a guard.

MIT — same as the original fablize. The upstream fablize by fivetaku is MIT licensed. This port adapts the design and procedures to the Hermes Agent ecosystem.