Loop Driven Engineering

Loop-Driven-Engineering

The Metaphor

The flight controller guiding four nested loops. LDD is Gradient Descent for Agents and this skill is the coordinator. The inner loop is the aircraft correcting pitch per second — direct edits to the code (∂L/∂code). The refinement loop is the approach pattern — polishing the landing without changing the aircraft (∂L/∂output). The outer loop is the training program that adjusts how pilots are taught — changing the skill itself (∂L/∂method). The CoT loop is the pilot's own step-by-step reasoning during the descent — each thought gated before it commits (∂L/∂thought). Each loop has its own budget, its own instruments, its own failure mode. The controller's first duty: name which loop is active. Confusion about which loop you're in is the canonical engineering error.

Overview

Engineering with an AI agent is gradient descent across four parameter spaces, not one. You plan, you try, you measure, you diagnose, you try again — on the code axis (inner loop, θ), the deliverable axis (refinement, y), the method axis (outer loop, m), and the reasoning-chain axis (CoT loop, t). This skill is the entry-point and coordinator: it picks which loop is active for the current task, dispatches the specialists, enforces the budget, and closes the loop only when regularizers hold.

The thinking-levels auto-dispatch runs before this skill and picks the rigor level (L0…L4) — that is the step-size scheduler for the whole optimizer. This skill runs inside the chosen level and orchestrates the actual descent.

Core principle: the loop closes when applicable gates are green and the doc-level mental model is current. Not when one test passes. Not when "it looks right." Not when the LLM sounds confident.

This skill is a flexible pattern, not a rigid procedure. Use judgment; adapt the structure to the task. But the budget, the escalation rule, and the loop-separation rule are hard.

See ../../docs/ldd/convergence.md for the four-loop model in full, ../../diagrams/four-axes-gradient-descent.svg for the top-level picture, and ../../diagrams/four-loops.svg for the loop-nesting view across all four axes.

When to Use

Invoke at the start of any non-trivial engineering task:

• Feature touching ≥2 files or ≥1 public surface
• Bugfix where the cause is not yet identified
• Refactor with observable behavior change (even if spec-preserving)
• Incident response
• Plan-driven work (a multi-step design under time pressure)
• Any work where you are tempted to "just start coding"

Do not use for:

• Trivial edits (rename a variable, fix a typo, delete dead code)
• Pure lookups or explanations
• Tasks already framed by a written plan you're executing verbatim

The Four Loops

LDD has four loops, each on a different parameter axis. Pick one at the start; running multiple simultaneously produces unstable gradients. See ../dialectical-cot/SKILL.md for the CoT loop (reasoning-chain axis, v0.8.0), which fires on verifiable multi-step reasoning tasks and is orthogonal to the three code-side loops below.

Inner Loop (θ = code, the default)

read context (CLAUDE.md / README / relevant docs — only what the task needs)
  ↓
plan (dialectical-reasoning → concrete next step)
  ↓
loop(k ≤ K_MAX = 5):
    observe failing signal
    ↓
    reproducibility-first (is this real signal or noise?)
    ↓
    e2e-driven-iteration: run E2E → measure loss_k → diagnose → fix → rerun
    ↓
    if green AND task-complete: break
    if failed: invoke root-cause-by-layer + loss-backprop-lens, next iteration
    ↓
    if k == K_MAX: STOP — escalate with the required shape, do not grind on
  ↓
close: docs-as-definition-of-done → commit

Refinement Loop (θ = deliverable)

Use when a deliverable exists, is "good enough but not great," and re-running from scratch would waste what worked. Dispatches iterative-refinement. Budget halves per iteration; stops on regression, plateau, or wall-time.

Outer Loop (θ = skills / rubrics / prompts)

Use when the same rubric violation recurs across 3+ distinct tasks. Dispatches method-evolution. Requires a task suite and Δloss_method measurement; rollback on regression.

Hard rule: if you can't say which loop you're in, stop. The wrong loop doesn't just waste budget — it can regress the artifact it was meant to improve.

The Budget: K_MAX = 5

Five iterations per task, default. If the loop has not closed after five, stop, write what failed and why, and escalate (to the user, to a colleague, to a different skill).

Why five: below that, you haven't explored; above it, you're flailing. Each extra iteration beyond five is a ~linear increase in sunk-cost bias and a ~geometric decrease in probability the next iteration will converge without a structural rethink.

Hard rule: when k hits K_MAX, produce the escalation deliverable below. It is a structured artifact, not a free-form apology. A 6th iteration without producing it is a process failure and is itself an outer-loop method-evolution trigger.

Escalation deliverable (v0.13.1 — concrete template):

1. What was tried, per iteration. One bullet per iteration, naming the skill that fired and the concrete change made (one line each). Example:

   • i1: reproducibility-first → confirmed fail is real (3/3 reruns)
   • i2: root-cause-by-layer → layer 3, contract violation on UserPayload.status (nullable vs. required)
   • i3: added defensive isinstance check — passed locally, sibling test regressed
   • i4: reverted; tightened schema instead — same regression
   • i5: swapped to discriminated union — 2 more siblings now fail

2. What kept failing — verbatim. Paste the exact error message / test name / gate output. No paraphrase. The reviewer needs to grep the artifact for this string.

3. Layer-4/5 diagnosis via root-cause-by-layer. Both layers must be named. Not "it's a design issue" — specify: "layer 4 (structural): UserValidator does two orthogonal jobs (schema check + behavior check) through one return type. layer 5 (conceptual): the domain has two invariants (shape, temporal order) but the code treats them as one."

4. One-sentence concrete architectural proposal. The sentence must name specific identifiers — file paths, class names, function names, data structures. Not "redesign module X", not "split the validator." Write the actual one-line design: "Split UserValidator into UserSchemaValidator (enforces R-rules 1–7) and UserBehaviorValidator (enforces R-rules 8–12); wire UserBehaviorValidator after UserSchemaValidator in the checkout pipeline; remove the dual-return-type entirely." If the sentence cannot be made concrete, the task is not ready for human escalation — it is ready for an architect-mode invocation first (the level scorer should have picked L3/L4 in that case).

5. Explicit ask. State what input you need, in one sentence: "Should I implement proposal (4) now, or do you want to see alternative architectures first via a /ldd-architect pass?" — not "let me know what you think."

Loss-backprop-lens is diagnostic, not designer. It tells you the step size was wrong (you were in a local minimum). It does not tell you what the bigger step is. That is the agent's job, in bullet (4). Escalations without a concrete bullet (4) get returned and the agent must re-do the K_MAX close before further work.

The Test Pyramid (cheap → expensive)

Run gates in order of cost. Only escalate when the current tier is green.

Anti-pattern: running E2E to catch a typo. It would have been caught by tier 1 in 2 seconds. Know your pyramid.

Anti-pattern: skipping tiers 1–3 because "E2E will catch everything." It won't, and when it fails you will have no idea at which layer it broke.

Sub-Skill Dispatch

This is a composition skill. Dispatch the right sub-skill at the right moment. Entries from this plugin are always available. Entries from other plugins (shown as external) are optional — if not installed, apply the principle inline.

You are not required to invoke all of them on every task. Pick what the loop actually needs at this step. But you are required to invoke the right one at the right moment — skipping reproducibility-first on a flake, or skipping docs-as-definition-of-done on a behavior change, is a process failure. When an external sub-skill is not available, apply its principle inline (draft the plan in the conversation rather than loading a plan-writing skill).

Red Flags — STOP, the loop is broken

• "I've tried this same fix 3 times, let me try a 4th variant" → local-minimum trap; invoke loss-backprop-lens, consider architectural step
• "The same pattern keeps happening across different tasks" → outer-loop signal; invoke method-evolution
• "I'll run E2E to see what breaks" (skipping tiers 1–3) → pyramid violation
• "I'll just commit and iterate in follow-ups" → budget discipline failure; write a plan
• "It failed / passed once, retry" → invoke reproducibility-first; one sample is not a gradient
• "This is iteration 6, one more should do it" → K_MAX violation; stop and escalate
• "Docs are fine for now, behavior change is subtle" → invoke docs-as-definition-of-done
• "I'll regenerate the doc from scratch instead of refining" → you conflated refinement with re-planning; invoke iterative-refinement
• "I'll edit the skill to match what just happened" → potential moving-target loss; invoke method-evolution with proper measurement
• "I don't need a plan, it's a small change" → "small" is the canonical rationalization for no-plan
• Same gate rejecting with the same reason ≥ 2 iterations → stagnation; abort, diagnose, restart with a fix at the structural origin
• No full drift scan in months on an active project → silent accumulation; invoke drift-detection

Escalation on Non-Convergence

When you can't close the loop, the productive response is not to try harder — it's to communicate. The right escalation has:

• What was tried — concrete, per iteration.
• What kept failing — the gate, test, or symptom, verbatim.
• The layer-4/5 diagnosis (via root-cause-by-layer).
• The step-size recommendation (via loss-backprop-lens): is this a local-fix problem that needs fresh eyes, or a structural problem that needs a different design?
• The explicit ask — what input do you need to proceed?

A well-escalated non-convergence beats a silently-limping sixth iteration every time.

How to Apply — checklist

1. Context read — only what the task needs. If the project has a CLAUDE.md / AGENTS.md, load it. Load the relevant docs. Do not read the whole repo.
2. Plan — one paragraph, run it through dialectical-reasoning if non-trivial. For multi-step work, write it out (use a plan-writing skill if available, otherwise draft the plan in the conversation).
3. Enter loop. k = 0, K_MAX = 5.
4. For each iteration: smallest coherent code change → fast gates → diagnose failures via sub-skills → decide step size.
5. Close — run docs-as-definition-of-done, verify completion with a verification skill if available, commit as one logical unit.
6. If K_MAX hit — escalate with the checklist above.

Counter-Examples — when this skill is overkill

• You're renaming a local variable. Just rename it.
• You're fixing a typo in a README. Fix it, commit.
• You're answering a factual lookup question. Read and answer.

The skill is for multi-step engineering under uncertainty, not for every interaction.

Real-World Shape

Good loop (feature: add --dry-run to myapp deploy):

1. Read cli.py, deploy.py, relevant README sections. (1 min)
2. Plan: dialectical pass — "thesis: add a flag, gate execution on it; antithesis: dry-run could leak partial state if deploy has side effects in plan-phase; synthesis: separate plan() from execute(), flag only skips execute()." (3 min)
3. k=1: refactor plan()/execute() split, unit tests for both. Tier-1+2 green. (15 min)
4. k=2: add flag, test dry-run path. Tier-1+2+3 green. (10 min)
5. Close: docs-as-definition-of-done — README L82 safety note, cli-reference table, user-guide preview section. (5 min)
6. Commit as one logical unit. Done in 2 iterations, budget to spare.

Bad loop (same task, no discipline):

1. Start coding the flag. k=1: works, but broke side-effect ordering. k=2: patch the ordering, broke a different test. k=3: patch that, broke idempotency. k=4: flaky. k=5: ship with retry. k=6: "one more fix." k=7: push through freeze window with TODO on README. Two weeks later, incident.

Related / Composed

In this plugin (always available):

• reproducibility-first — gate before any gradient use
• root-cause-by-layer — the 5-layer debugging ladder
• loss-backprop-lens — edit-size calibration and signal-vs-noise
• e2e-driven-iteration — measure-every-iteration inner-loop rhythm
• dialectical-reasoning — planning and recommendations
• iterative-refinement — y-axis (deliverable) optimization
• method-evolution — θ-axis (skill / rubric) outer-loop optimization
• drift-detection — periodic full-repo scan for cumulative drift
• docs-as-definition-of-done — the close of the loop

Optional companions (Claude Code superpowers plugin, Codex built-ins, or equivalent):

• brainstorming — open-ended problem framing
• writing-plans — externalized multi-step plans
• test-driven-development — red-green discipline when writing code
• verification-before-completion — evidence before success claims
• systematic-debugging — overlaps with root-cause-by-layer; prefer the 5-layer ladder for the explicit discipline, the broader investigation framing when you need to find the bug rather than diagnose one already localized