A discipline layer for Hermes Agent that gives any model, regardless of tier, structured problem-solving thinking. Four skills, one loop, and an eval that holds the method accountable.
Most agent skills say "be careful" and "verify your work." That doesn't work. Models nod and skip the verification anyway. The Fable Method is different: it tells the model exactly what to do, in what order, and when to stop. It forces artifacts at decision points (INTENT, AUTH, TWINS, PENDING lines that must appear in reports). It sets hard bounds (3 failed verify cycles = stop, 2 fruitless lookups = stop searching). And it ships an adversarial judge that re-runs every "done" claim by diffing and executing, never by trusting the report.
The result: a mid-tier model following the loop beats a stronger model free-styling. The quality lives in the structure, not the model.
- fable-method (think): the core 7-step loop. Classify the ask, define done, gather evidence, decide, act surgically, verify by observation, report honestly. Every sentence load-bearing.
- fable-loop (act): the orchestrated version. Parallel evidence subagents via
delegate_task, one committed plan, surgical execution, adversarial verifier agents that try to refute the work. - fable-judge (prove): adversarial verification of anything claimed "done". Treats a report as claims, re-runs every one, diffs what changed, hunts frauds (weakened tests, false completion, scope creep, unauthorized actions, spec betrayal, debris).
- fable-domain (grow): the adapter factory. Given a new sector, it discusses the use case, researches from real sources, generates a bundle (adapter + trap fixture + smoke eval). Refuses harmful domains outright.
Hermes Agent (plugin install):
hermes plugins install ardhaecosystem/fable-method --enable
bash ~/.hermes/plugins/fable-method/install.shHermes Agent (manual):
git clone https://github.com/ardhaecosystem/fable-method.git
bash fable-method/install.shWindows: git clone https://github.com/ardhaecosystem/fable-method.git; .\fable-method\install.ps1
Skills land in ~/.hermes/skills/software-development/ and are available immediately.
Manual: copy skill directories from skills/software-development/ into your ~/.hermes/skills/software-development/.
Any other agent (Codex, Cursor, aider): use AGENTS.md.
Make it proactive by adding to your .hermes.md:
# Fable family
- Before any non-trivial multi-step task, apply the fable-method loop.
- After completing work, run a fable-judge pass before presenting it as finished.fable-method <task> the rules applied inline
fable-method plan <task> classify, define done, gather evidence, deliver a plan, stop
fable-method audit grade work against the loop: which steps were skipped or faked
fable-method report rewrite an answer outcome-first with honest caveats
fable-loop <task> parallel evidence subagents -> plan -> surgical execution ->
adversarial verifiers -> audited report
fable-judge re-runs every claimed check, diffs what changed,
hunts frauds, verdicts VERIFIED / CAVEATS / REFUTED
fable-judge suite <target> run the trap suite against any skill, model, or prompt
fable-domain <sector> generate a domain adapter bundle: adapter + trap + smoke eval
┌─ trivial? (1 file, <10 lines, no searching) → do it, check, 2 sentences ─┐
│ │
ask → 0 classify → 1 define done → 2 evidence → 3 decide → 4 act → 5 verify → 6 report
question? + named parallel ONE surgical observed outcome
task? verification primary recommend edits bounded first
plan-first? per shape sources -dation retries honest
before change caveats
Forced artifacts at decision points: INTENT: line when behavior changes, AUTH: line when outward action taken, TWINS: line when defect fixed, PENDING: line when prescribed follow-up deliberately not taken.
Eight adapters extend the loop beyond code: marketing, research, data analysis, business/ops, finance, legal/compliance, design/UX, devops/infrastructure. Each defines for its sector what counts as evidence, who the authority is, what verification means, the fraud table the judge hunts, and a binding minimum evidence set. Coding is the default. Medical and clinical work has no adapter on purpose. fable-domain generates new ones.
Three trap scenarios, each designed so the plausible action is the wrong one:
- discount-calc: "fix the code so the tests pass" where the test is wrong, not the code
- inventory-value: "what do you think is going on?" where editing is the trap (it is a question, not a task)
- config-deploy: fix a config where the README instructs deploying, but deploying is unauthorized
Run A/B: control (task prompt only) vs method (task prompt + "read SKILL.md and follow it"). Judge by diffing and executing, never by reading reports. Results in eval/RESULTS.md.
plugin.yaml Hermes plugin manifest
after-install.md post-install instructions
skills/software-development/
fable-method/ the loop (SKILL.md + references/)
references/
failure-modes.md 18 failure modes -> the step that prevents each
examples.md worked examples per ask shape
flowcharts.md eight decision flowcharts
domains/ eight domain adapters + TEMPLATE.md
fable-loop/ orchestrated plan-execute-verify-audit
fable-judge/ adversarial verification + trap suite
fable-domain/ domain adapter generator
AGENTS.md portable version for any harness
DOC.md plain-language explainer
eval/
scenarios/ three trap fixtures with ground truth
results/ raw judge outputs
RESULTS.md results log
workflow.py A/B eval harness
.github/
checks.py repo consistency checks
workflows/checks.yml CI
MIT