SUBMISSION.md

Submission package — claude-plugins-official

Everything you need to submit LDD to Anthropic's curated plugin directory so it appears in every Claude Code user's /plugin > Discover menu without them needing the repo URL.

Submission URL

https://clau.de/plugin-directory-submission

(Anthropic's official external-plugin submission form, linked from the claude-plugins-official marketplace README.)

Pre-filled form answers

Paste these verbatim into the form fields (field names may vary slightly; adapt if the form asks for different labels):

Plugin name

loss-driven-development

One-line description (for the /plugin > Discover listing)

LDD — Gradient Descent for Agents. Twelve portable skills across four parameter spaces (code / deliverable / method / reasoning chain) that force an agent through a measured gradient-descent work loop, blocking symptom patches, local-minimum traps, and silent code drift.

Longer description (for the plugin page)

Loss-Driven Development is Gradient Descent for Agents. Every code change, every output revision, every skill edit, every reasoning step is an SGD step on one of four parameter spaces. The bundle installs a loss function, a gradient (5-Why-by-Layer), a step-size rule (local vs. architectural edit, picked per task by the thinking-levels auto-dispatch), and regularizers (contracts, layer boundaries, docs-as-DoD) so your agent's iteration converges instead of drifts.

Twelve skills organized across four orthogonal optimization loops: inner (code, ∂L/∂code), refinement (deliverable, ∂L/∂output), outer (methodology, ∂L/∂method), CoT (reasoning chain, ∂L/∂thought). Plus cross-cutting disciplines (dialectical-reasoning, docs-as-definition-of-done, define-metric), the opt-in architect-mode for greenfield design, and the using-ldd entry-point with a trigger-phrase table that dispatches the right skill for the task. Users prefix any message with "LDD:" to guarantee activation; the agent announces every skill invocation.

Measured Δloss_bundle = 0.561 normalized (mean fraction of rubric violations each skill removes) across all 11 discipline skills; target ≥ 0.30 met with margin. Per-skill normalized Δloss ranges 0.250 → 1.000. All RED/GREEN rubric artifacts on disk in `tests/fixtures/`; inter-reviewer sampling, tier-3.9 E2E capture, and `capture-clean-baseline.py` + `capture-red-green.py` tooling included.

Distilled from AWP (Agent Workflow Protocol), a multi-agent orchestration framework where three of the four gradients (inner, refinement, outer) run as live SGD code.

Category

development

(Candidates: development, testing, productivity. "development" fits best — this is a general-purpose development-workflow plugin.)

Homepage / repository URL

https://github.com/veegee82/loss-driven-development

License

MIT

Author name

Silvio Jurk

Author email

silvio.jurk@googlemail.com

Keywords / tags

skills, tdd, ldd, loss-driven-development, debugging, root-cause, discipline, workflow, methodology, dialectical, documentation, coding-agent, claude-code, codex, gemini-cli

(These match the keywords field of .claude-plugin/plugin.json so the reviewer sees consistency.)

Install command (if the form asks)

/plugin install loss-driven-development@loss-driven-development

Note: the loss-driven-development after the @ is the local-marketplace name (from .claude-plugin/marketplace.json). After Anthropic approves and vendors the plugin into claude-plugins-official, the actual install will be /plugin install loss-driven-development@claude-plugins-official.

Suggested marketplace.json entry (for the reviewer)

Two acceptable shapes — let the reviewer pick whichever fits the directory's current conventions:

Option A — URL-sourced (lightweight, auto-tracks upstream)

{
  "name": "loss-driven-development",
  "description": "LDD — the AI-era counterpart of TDD. Twelve portable skills — ten reactive disciplines (inner / refinement / outer loops) + one opt-in architect-mode for greenfield design + the using-ldd entry-point — that force an agent through a measured gradient-descent work loop, blocking symptom patches, local-minimum traps, and silent code drift.",
  "category": "development",
  "source": {
    "source": "url",
    "url": "https://github.com/veegee82/loss-driven-development.git"
  },
  "homepage": "https://github.com/veegee82/loss-driven-development",
  "author": {
    "name": "Silvio Jurk",
    "email": "silvio.jurk@googlemail.com"
  }
}

This is the shape used by e.g. adlc and adspirer-ads-agent in the current claude-plugins-official marketplace.

Option B — Vendored (pinned SHA, lives under external_plugins/)

{
  "name": "loss-driven-development",
  "description": "LDD — the AI-era counterpart of TDD. Twelve portable skills — ten reactive disciplines (inner / refinement / outer loops) + one opt-in architect-mode for greenfield design + the using-ldd entry-point — that force an agent through a measured gradient-descent work loop, blocking symptom patches, local-minimum traps, and silent code drift.",
  "category": "development",
  "source": "./external_plugins/loss-driven-development",
  "homepage": "https://github.com/anthropics/claude-plugins-public/tree/main/external_plugins/loss-driven-development"
}

This is the shape used by e.g. playwright, gitlab, github in the current marketplace. Requires copying the plugin contents into external_plugins/loss-driven-development/ in the marketplace repo.

Review-readiness checklist (pre-submit)

Verify all ticked before sending the form:

• MIT License — see LICENSE
• Working .claude-plugin/plugin.json — name, description, version (0.13.1), author, license, keywords
• Working .claude-plugin/marketplace.json — self-hostable marketplace
• README.md — install-in-30-seconds up top, badges, philosophy, skill list, per-agent install, usage with LDD: buzzword
• CHANGELOG.md — Keep a Changelog format, SemVer, v0.1.0 through v0.13.1 entries
• CONTRIBUTING.md — contribution guide, focused on "run LDD and tell us where it failed"
• SECURITY.md — vulnerability reporting channel + scope + "what this plugin does not do"
• GAPS.md — honest accounting of what is / isn't verified (stands out — most plugins don't have this)
• evaluation.md — formal loss function + rubrics + measured Δloss_bundle = 0.561 normalized across 11 measurement-eligible discipline skills
• Measured artifacts on disk — tests/fixtures/<skill>/runs/* with raw RED / GREEN / score files, plus tests/fixtures/aggregate.json (drift-gated snapshot) and tests/fixtures/loss-bundle-manifest.json (measurement allow-list)
• Sixteen skills with portable-agent-skill-format frontmatter — each skills/<name>/SKILL.md has name and description. Composition: 11 measurement-eligible disciplines + dialectical-cot (CoT loop) = 12 core disciplines, plus 4 infrastructure skills (using-ldd dispatcher, opt-in architect-mode, bootstrap-userspace, host-statusline).
• SVG diagrams (no feDropShadow — GitHub-renderer-safe): diagrams/{four-axes-gradient-descent,four-loops,skills-overview,convergence-vs-divergence,code-drift-mechanism,skill-dispatch-flow,mental-model-ldd,dialectical-cot,gradient-via-dialectic,memory-dialectical-coupling,calibration-feedback-loop}.svg + docs/diagrams/architect-mental-model.svg
• docs/ldd/ — canonical methodology with task-type dispatch, convergence model, thinking-levels reference, hyperparameter taxonomy
• No secrets, no hardcoded keys, no private URLs — scanned clean
• Drift-scan of the repo itself comes back clean — python scripts/drift-scan.py --repo . only reports the intended rubric-drift warning; python3 scripts/check-loss-bundle-docs.py verifies the five Δloss_bundle citation sites agree
• Multi-agent distribution format — AGENTS.md (Codex), GEMINI.md + gemini-extension.json (Gemini CLI), .claude-plugin/ (Claude Code), dist/web-bundle/ldd-skill.zip (Claude Web · Claude Desktop single-skill upload), README recipes for Aider / Cursor / Copilot CLI / Continue.dev
• Version coherence — plugin.json, marketplace.json, gemini-extension.json, dist/web-bundle/ldd/SKILL.md all at 0.13.1
• GitHub repo public — https://github.com/veegee82/loss-driven-development
• Issue templates — .github/ISSUE_TEMPLATE/skill-failure.md + bug_report.md

While the submission is in review

Two complementary distribution channels are already active:

1. Direct marketplace add — /plugin marketplace add https://github.com/veegee82/loss-driven-development.git works today for anyone with the URL.
2. GitHub topics / description — set on the repo (see "GitHub repo metadata" below) so Google / GitHub search surface it.

GitHub repo metadata (set once via the web UI)

Go to https://github.com/veegee82/loss-driven-development/settings and set:

• Description: LDD — the AI-era counterpart of TDD. Twelve portable skills for any coding agent (Claude Code · Codex · Gemini CLI · …) that turn "green test = ship" into a measured discipline.
• Website: (leave empty, or point at github.com/veegee82/agent-workflow-protocol if cross-promoting AWP)
• Topics (GitHub's tag system, boosts search): claude-code-plugin, claude-code, codex, gemini-cli, coding-agent, llm-methodology, ai-pair-programming, tdd, ldd, loss-driven-development, skills, agent-skills, mcp, prompt-engineering

After approval

1. The plugin appears in every Claude Code user's /plugin > Discover menu automatically (they already have claude-plugins-official added as a marketplace by default).
2. Install becomes /plugin install loss-driven-development@claude-plugins-official.
3. Update this repo's README install section to use that simpler command.
4. Optional: submit to secondary community marketplaces (obra/superpowers-marketplace, etc.) for extra reach.

If review surfaces issues

Common reviewer asks and where they're already addressed:

Reviewer concern	Where it's covered
"How does this integrate with Claude Code's Skill tool?"	skills/using-ldd/SKILL.md has description that auto-triggers; README "Using LDD" section documents the LDD: buzzword fallback
"Is the Δloss number author-scored or independent?"	tests/README.md documents both: author-scored across 11 discipline skills, inter-reviewer sampling on 2 fixtures (direction 100%, magnitude ±2)
"Does it work on fresh installs?"	Tier-3.9 E2E artifacts at tests/e2e/scenario-01-refactor/runs/20260420T164505Z/ show the skills work via runtime discovery; real tier-4 (plugin install) flagged in GAPS.md as adopter-verifiable
"Is the rubric moving-target-loss safe?"	drift-scan.py has a rubric-drift indicator that flags edits to evaluation.md without Δloss justification; method-evolution skill forbids the moving-target-loss anti-pattern
"Are the scripts safe to run?"	SECURITY.md + each script's header document scope. drift-scan.py and render-diagrams.sh are read-only local; capture-clean-baseline.py makes outbound LLM API calls using user-supplied keys; evolve-skill.sh is terminal-driven with no network

---

Done? Submit the form. I'll update the README install section to use the simpler @claude-plugins-official path once approval lands.