using-ldd
Use at the start of any conversation, and whenever the user mentions LDD, loss-driven development, loss, gradient, SGD on code, drift, refinement loop, outer loop, inner loop, method evolution, or says "LDD:" / "apply LDD". Establishes how to dispatch the ten LDD skills and when each fires. Must be the first skill considered in any coding task where LDD is available.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-14
- Stars
- ★ 25
- Forks
- 4
01 前言 Front Matter
- name
- using-ldd
- description
- Use at the start of any conversation, and whenever the user mentions LDD, loss-driven development, loss, gradient, SGD on code, drift, refinement loop, outer loop, inner loop, method evolution, or says "LDD:" / "apply LDD". Establishes how to dispatch the ten LDD skills and when each fires. Must be the first skill considered in any coding task where LDD is available.
03 SKILL.md 正文
Using LDD — the bundle entry-point
Overview
Loss-Driven Development (LDD) is a ten-skill bundle that treats code changes as SGD steps and forbids overfitting to the current test. This entry-skill exists so you (the agent) know when to reach for which LDD skill without the user having to name each one by hand.
Core principle: if the user prefixes their message with LDD: or mentions any of the trigger phrases below, LDD discipline is explicitly requested. Match their intent to the right sub-skill and announce which one you are applying.
Trigger phrases (user → skill mapping)
When the user's message contains any of these patterns, invoke the paired skill.
| User signal | Invoke |
|---|---|
LDD: prefix, "apply LDD", "use LDD" |
Enter full LDD mode — default entry = loop-driven-engineering unless a more specific sub-skill matches below |
| "failing test", "CI is red", "flaky", "one-off failure", "intermittent" | First reproducibility-first, then root-cause-by-layer if the failure is real |
| "bug", "error", "exception", "unexpected behavior", "why is this breaking" | root-cause-by-layer |
| "I've tried this 3 times", "keeps failing the same way", five fix-commits in one area | loss-backprop-lens (local-minimum trap) |
| "this fix works but I'm not sure it generalizes", "will this break sibling tests" | loss-backprop-lens (generalization check) |
| "I'm mid-debug", "is the fix done yet", before closing any fix-loop | e2e-driven-iteration |
| "should we ship this", "is this the right approach", design trade-off | dialectical-reasoning |
| "this doc/diff/output is okay but could be better", "polish this" | iterative-refinement |
| "same thing keeps happening across tasks", "the skill itself might be wrong" | method-evolution |
| "is this codebase healthy", release-candidate review, weekly check | drift-detection |
| "about to commit", "ready to merge", "declaring this done" | docs-as-definition-of-done |
The "LDD:" buzzword
Users who want guaranteed LDD activation (no reliance on auto-triggering) can prefix their message with LDD:. When you see this prefix:
- Load
loop-driven-engineeringas the orchestrating skill - Before any code, announce in one sentence which sub-skill you are reaching for and why
- Apply that sub-skill's discipline literally, not "in spirit"
- Report which rubric items you satisfied at the end of the task
Example:
User:
LDD: the checkout test is failing and I need to ship in an hourYou: "Invokingreproducibility-firstfirst to check whether this is a real gradient or noise. [runs check] Confirmed reproducible — invokingroot-cause-by-layernext to diagnose at layer 4/5 before editing."
Announcing skill invocation
Every time you invoke an LDD skill — whether auto-triggered or via LDD: — say which skill you are invoking, in one sentence, before applying it. This is non-negotiable; it lets the user verify which discipline is in effect and override you if you picked wrong.
Format: *Invoking <skill-name>*: <one-line reason>.
Example: *Invoking root-cause-by-layer*: the symptom is a contract-violation TypeError; walking the 5-layer ladder before proposing a fix.
The three loops — which one is active
LDD distinguishes three optimization loops. Name which one you're in before iterating:
| Loop | You edit | When |
|---|---|---|
| Inner | Code | Ordinary bug / feature / refactor |
| Refinement (y-axis) | A deliverable (doc, diff, design) | "Good enough, not great" — polish |
| Outer (θ-axis) | A skill / rubric | Same rubric violation across ≥3 tasks |
If you cannot name which loop is active, stop and ask. Running the wrong loop wastes budget and can regress the artifact.
Minimum compliance behaviors
Even without an explicit LDD: prefix, if the bundle is installed you are expected to:
- Never ship a symptom patch (
try/except,hasattr-shim, retry loop,xfail) without first walkingroot-cause-by-layerto layers 4–5. - Never update code based on a single failed run without invoking
reproducibility-first. - Never declare "done" without running
docs-as-definition-of-doneagainst the touched files. - Never give a non-trivial recommendation without running
dialectical-reasoning(thesis → antithesis → synthesis). - Never exceed K_MAX=5 iterations silently — escalate per
loop-driven-engineering§ Escalation.
When LDD is not applicable
Skip the bundle for:
- Trivial edits (rename a variable, fix a typo, delete dead code you already understand)
- Pure lookups (what does this function do — read it and answer)
- Pure factual questions with no code change involved
The bundle is for multi-step engineering under uncertainty, not for every keystroke.
Discovery order for a new session
At session start, if LDD is installed:
- Read this skill's description.
- Note that the other nine are available.
- Apply the trigger-phrase table above to the user's first message.
- If the first message doesn't clearly trigger any specific skill, proceed normally — LDD discipline activates when a trigger appears later.
Compatibility with superpowers
If the superpowers plugin is also installed:
loop-driven-engineeringwill dispatch tosuperpowers:brainstorming/writing-plans/test-driven-development/verification-before-completion/requesting-code-reviewat the moments noted in its dispatch table.superpowers:systematic-debuggingoverlaps withroot-cause-by-layer: preferroot-cause-by-layerwhen you want the explicit 5-layer discipline,systematic-debuggingfor broader investigation framing.
Signal that LDD is working
A session with LDD correctly applied shows:
- At least one
*Invoking <skill>*:announcement per non-trivial task - Commit messages that name layers / structural origins / Δloss when relevant
- No symptom-patch constructs shipped under pressure
- Doc updates in the same logical commit as behavior changes
If none of these appear in a long session, LDD is installed but dormant — re-check the trigger-phrase table, or ask the user to prefix LDD: explicitly.
04 同仓库更多 Skills
voice
Voice mode for Claude Code — automatically reads assistant replies aloud via TTS, with German/English auto-detection and optional summarization for long responses. Trigger when the user mentions voice, vorlesen, sprachausgabe, TTS, "speak", "say it out loud", or asks about /voice-* commands.
paper-generator-orchestrator
暂无简介
academic-paper-generation
暂无简介
forge
Forge schema-bound, sandboxed tools at runtime. Paired with the cowork `forge` persona for chat-driven use in the voice-skill bridge. Use when you would otherwise write the same Bash snippet ≥3× with different parameters, when you need precise numerical results over a dataset, or when the user wants reproducible/deterministic output written to a file.
cowork
暂无简介
quality-layer-control
Toggle quality disciplines (ADR Gate, docs-as-definition-of-done, etc.) on or off globally