fable-discipline MCP Server

fable-discipline

Reusable workflow discipline for Claude Code agents.

fable-discipline is a Claude Code plugin that makes agentic software work follow repeatable working patterns: design before code, verify after edits, separate author from reviewer, preserve verified state between sessions, and report uncertainty honestly.

It is not a model upgrade and not a magic prompt pack. It shapes procedure, not raw capability — a checklist, not a capability transplant. It can't raise a model's reasoning ceiling.

It runs on a three-layer model:

• The plugin — the generic HOW: how to plan, build, verify, review, remember, and report.
• Your repo docs — the local WHAT: CLAUDE.md, DESIGN.md, STATE.md hold the product truth.
• Your prompt — the current goal: e.g. "Review the new onboarding flow."

Install

/plugin marketplace add petrkindlmann/fable-discipline
/plugin install fable-discipline@kindlmann-workflows

Or add the marketplace from a local path or git URL, then install.

How users actually use this

You do not read the skills manually before every session. Install the plugin, then work normally in Claude Code. The relevant skill auto-applies based on the situation.

Example prompts:

Plan the next billing milestone.
Build the analytics dashboard redesign.
Review the v1.8 changes before I merge.
Resume where we left off and check the project state first.
Build a CLI that scans Playwright tests for flaky selectors.

For serious product work, use the explicit milestone commands:

/milestone-research checkout-redesign
/milestone-build checkout-redesign
/milestone-review checkout-redesign

Use them in order. Approve the design doc between research and build.

Which command should I run?

Situation	What to do
You are not sure what to build yet	/milestone-research name
You have an approved design and want implementation	/milestone-build name
You want a review before merge, release, or deploy	/milestone-review name
You are building a CLI, plugin, MCP server, library, SDK, or skills pack	Describe the artifact. artifact-build should trigger.
You are redesigning UI or using Stitch/Figma output	Describe the redesign. design-workflow should trigger.
You are running audits, scans, benchmarks, or recurring reports	Describe the measurement job. measurement-pipeline should trigger.
You are testing a hypothesis or writing an evidence-backed report	Describe the investigation. research-investigation should trigger.
You are delegating to subagents or running loops	Describe the fan-out. agent-orchestration should trigger.
You are resuming old work	Ask Claude to read state first. compounding-memory should trigger.

The three main flows

1. Ship a product milestone

Use this for feature work inside an existing app.

/milestone-research checkout-redesign

Expected behavior: Claude reads the repo's CLAUDE.md, researches the feature, checks constraints, and writes a scoped design doc. It should not start coding yet.

/milestone-build checkout-redesign

Expected behavior: Claude builds from the approved design in atomic, verified steps. It runs the repo's smallest meaningful checks after each step.

/milestone-review checkout-redesign

Expected behavior: Claude runs an adversarial review, classifies findings honestly, and fixes them in numbered batches.

2. Build a standalone artifact

Use this for CLIs, MCP servers, plugins, libraries, SDKs, and skills packs.

Example:

Build a new CLI that scans Playwright tests and reports flaky selectors. Start with a design spec, then a checkbox implementation plan, then build it with tests.

Expected behavior: design spec first, implementation plan second, modular TDD build, review pass, live validation, release packaging.

3. Redesign product UI

Use this when creating a design system, writing a DESIGN.md contract, generating screens in Stitch/Figma, or porting generated UI to code.

Example:

Create a DESIGN.md contract for the analytics dashboard, then port the generated Stitch screen into real components. Verify the implementation against both the screen and the contract.

Expected behavior: read the full design contract, port screen-by-screen, avoid fake numbers and generic AI UI clichés, and verify against the real viewport.

Skills

Each skill is small, composable, and situation-triggered. Shared detail lives in skills/<name>/references/ so the same instructions are not duplicated.

Building and shipping

Skill	User-facing label	Fires when	Discipline
milestone-workflow	Ship a milestone	Shipping a feature inside an existing app	Research and design, then build, then adversarial review. Also powers /milestone-research, /milestone-build, /milestone-review.
artifact-build	Build an artifact	Building a CLI, MCP server, library, plugin, SDK, or skills pack	Design spec, checkbox TDD plan, modular build, review pass, live validation, release packaging.
design-workflow	Design system workflow	Building or redesigning product UI	DESIGN.md contract, generated screens ported to code, verification against the contract and rendered screen.

Investigating and measuring

Skill	User-facing label	Fires when	Discipline
research-investigation	Research with controls	Empirical analysis, forensics, reverse engineering, evidence-backed reports	Pre-registration, positive and negative controls, calibrated negatives, source-engagement ledger, capability-bound caveats.
measurement-pipeline	Measurement audit	Recurring audits, benchmark matrices, scans, dashboards	Complete gap accounting, evidence as proof, infra-vs-real failure classification, one-command recurring run, stakeholder report.

Always-on support

Skill	User-facing label	Fires when	Discipline
agent-orchestration	Agent orchestration	Multi-agent, fan-out, delegation, loops	Model-tier routing, verifier is not author, rubric-gated loops, dynamic workflows, worktree isolation.
compounding-memory	Project memory	STATE.md, MEMORY.md, resume work, persistence	Fail, investigate, verify, distill, consult. Write before leaving, read at start, tag facts with Verified-by:.
execution-rules	Execution discipline	Any hands-on coding or agent task	Literal instruction-following, thinking calibration, sufficient context, act then verify, no AI slop.

Docs

• How to use this
• Copy-paste examples
• Troubleshooting — when the agent goes off the rails

Suggested repo setup

The plugin works best when each project has a clear CLAUDE.md:

# CLAUDE.md

## Verify commands
- npm run lint
- npm run typecheck
- npm test

## Deploy model
Describe what deploys on push, what requires approval, and what must not be touched casually.

## Project-specific rules
List the product's moat, data honesty rules, commit style, and file locations.

## State files
- .planning/STATE.md
- .planning/ROADMAP.md

The plugin provides the working discipline. The repo provides commands, constraints, and product truth.

Contributing

The collection grows by distilling repeatable workflow patterns out of real agentic work. Contribute patterns of work, never proprietary text.

See CONTRIBUTING.md for the method and the ethical line. See PROVENANCE.md for source attribution.