prompt-optimization
A single-plugin Claude Code marketplace hosting the prompt-optimization plugin: a phased workflow that turns a draft prompt, system message, CLAUDE.md, or multi-agent orchestrator into production-grade instructions for Claude Fable 5 (with no-silent-fallback deployment guidance for the Fable 5 classifier / prior-flagship fallback architecture).
Install
Claude Code (CLI)
claude plugin marketplace add ZaBrisket/prompt-optimization
claude plugin install prompt-optimization@prompt-optimization
Or, from within an interactive Claude Code session:
/plugin marketplace add ZaBrisket/prompt-optimization
/plugin install prompt-optimization@prompt-optimization
The first command registers this repo as a marketplace; the second installs the plugin from it (<plugin>@<marketplace>). Restart Claude Code if prompted so the new skill and agents load.
Dynamic-workflow orchestration is a research preview and requires Claude Code v2.1.154 or later.
Cowork (Desktop)
- In the Cowork desktop app, open the Customize menu and add this repo (
ZaBrisket/prompt-optimization) as a marketplace.
- Install prompt-optimization from the marketplace.
Local development
claude plugin marketplace add /path/to/this/repo
claude plugin install prompt-optimization@prompt-optimization
What you get
- Skill —
prompt-optimization — auto-triggers when you provide a draft prompt to improve, ask to optimize/harden/refine a prompt, or describe a multi-agent research workflow. Runs a sequential protocol: intent extraction, complexity triage, widget-based clarification, mandatory live calibration against Anthropic documentation, landscape research, draft analysis, prompt construction, delta analysis, and a final QC pass with a clean file write. The draft-analysis anti-pattern sweep is calibrated against the Fable 5 / Mythos 5 system card (six-pattern long-horizon failure taxonomy, fallback semantics, refusal inversion, mixed prompt-injection picture, multi-agent harness patterns). Invoke explicitly with /prompt-optimization if the auto-trigger doesn't fire.
- Agent —
landscape-research — a read-only research worker dispatched as part of the orchestrated Phase 2 (mode 2-O) dynamic workflow — one lane per sub-domain or query-taxonomy cluster — with a parallel-subagent fallback when the workflow runtime is unavailable; results synthesize centrally.
- Agent —
devils-advocate — a read-only adversarial / confirmatory worker dispatched by the verify-and-converge stage of every orchestrated-research deliverable: adversarial mode for thesis-advancing deliverables, confirmatory mode for purely descriptive inventory / fact-extract deliverables.
Orchestrated landscape research (mode 2-O)
When the draft is multi-sub-domain, multi-stakeholder, or contested, Phase 2 fans out via a Claude Code dynamic workflow: one landscape-research lane per sub-domain or query-taxonomy cluster, plus a dedicated adversarial lane for the verify-and-converge stage, each searching deeply in its own isolated context. The orchestrator's reduce step collects the lane returns and runs synthesis and the neutrality / coverage-bias gates on the aggregate corpus. Fan-out is bounded by ~6 lanes (synthesis bandwidth), the runtime caps (≤16 concurrent / ≤1,000 total agents per run), per-lane search budgets, and at most one remediation pass. When the dynamic-workflow runtime is unavailable, Phase 2 falls back to parallel subagents via the Agent/Task tool (or single-threaded execution on claude.ai chat / bare API) automatically.
The orchestrated-research deliverable the plugin produces is itself a CLAUDE.md that orchestrates a dynamic workflow.
Repository layout
.
├── README.md # this file
├── .claude-plugin/marketplace.json # marketplace definition
└── plugins/
└── prompt-optimization/
├── .claude-plugin/plugin.json # plugin manifest
├── skills/prompt-optimization/ # SKILL.md + references/
├── agents/landscape-research.md # landscape research worker
├── agents/devils-advocate.md # devil's-advocate / confirmatory worker
└── README.md
Author
Mac Zabriskie
