Skip to content

getting-started.md

Latest commit

 

History

History
76 lines (48 loc) · 3.31 KB

getting-started.md

File metadata and controls

76 lines (48 loc) · 3.31 KB
Raw

Getting started with LDD

One page. Read this if (a) you're a user about to install LDD, or (b) you're an agent encountering LDD for the first time.

For the user

After install, try the bundle on a real task:

Option 1 — explicit activation (recommended first time)

Prefix your message with LDD::

LDD: the checkout test is failing and I need to ship in an hour

The agent will announce which skill it is invoking:

Invoking reproducibility-first: one failure is one sample; checking reproducibility before treating this as a gradient.

If you do not see an *Invoking <skill>*: line, LDD is installed but dormant — re-check install per ../../README.md#verifying-ldd-is-active.

Option 2 — trust auto-trigger

Just describe your task naturally:

I've tried fixing this flaky test three times and it keeps breaking different tests

The agent should auto-invoke loss-backprop-lens (local-minimum trap trigger). If it doesn't, fall back to Option 1.

Option 3 — name the skill

Use root-cause-by-layer on this TypeError

Works in Claude Code (/loss-driven-development:root-cause-by-layer), Codex (names in AGENTS.md), Gemini CLI (GEMINI.md imports).

For the agent

You are in a project that has LDD available. On every non-trivial task:

  1. Check task-types.md — pick the task type matching the user's message.
  2. Load only the MDs that task type requires. Do not pre-load the whole methodology.
  3. Announce invocations — *Invoking <skill-name>*: <one-line reason>. before applying each skill.
  4. Respect minimum-compliance behaviors (from ../../skills/using-ldd/SKILL.md):
    • No symptom patches without 5-layer diagnosis
    • No code edit on single-sample failure without reproduction
    • No "done" without docs synced
    • No non-trivial recommendation without dialectical pass
    • No silent iteration 6

The shortest possible summary

LDD = Gradient Descent for Agents. SGD across four parameter spaces — code (inner), deliverable (refinement), method (outer), reasoning chain (CoT). If your edit is not a real gradient step on a named axis (reproducible signal + honest diagnosis + right step size + regularizers intact), it's noise injected into that parameter space.

That sentence is the whole bundle. Everything else — the four loops, the twelve skills, the thinking-levels auto-dispatch, the trace-block format — is scaffolding to keep the sentence honest under pressure. Long-form theory: ../theory.md.

Five-minute installed-and-working checklist

You'll know LDD is working when:

  • *Invoking <skill>*: lines appear in agent responses to non-trivial tasks
  • Commit messages name layers, Δloss, or structural origins when relevant
  • Agent pushes back on "just rerun" / "try/except wrapper" / "I'll update docs later"
  • Design recommendations include explicit thesis / antithesis / synthesis

None of these happen? Check:

  • Is the plugin actually installed? (/plugin list in Claude Code)
  • Is the skill directory discoverable? (see install per your agent)
  • Did you try the LDD: prefix?

Still not working? Open an issue with .github/ISSUE_TEMPLATE/skill-failure.md.