AUTO-CODING ANALYSIS // v2.0
Engineering Leadership Briefing
Auto-Coding
Reality Check

LLM-assisted dev promises 10× delivery. But faster code isn't faster software — and the hidden costs compound fast.

By Patrick Delaney / pwdel github.com/pwdel
01The Business Problem — does the paradigm even work?
0214 Observed Problems in Production LLM Systems
03Hard Interventions — Hooks & Deterministic Gates
04Soft Interventions — Prompts, Context & Workflow
05Quick-Fire Quiz: Know Your Pitfalls?
06Claude Code Best Practices
07ROI Verdict & Recommendations
SLIDE 01 / BUSINESS PROBLEM
01 — The Business Problem

Who Gives A Shit If
You Code Faster?

Unsupervised code is a ticking time bomb set to destroy value at a future date.

Risk Factors
Low risk profile. At these settings, LLM assistance is net positive. Move the sliders to see how fast that changes.
SAFE
LOW COSTCATASTROPHIC
future cost ∝ eagent · complexity · risk
SLIDE 02 / OBSERVED PROBLEMS
02 — Observed Problems in LLM Coding Systems

14 Ways It Goes Wrong

←
Select a problem
SLIDE 03 / HARD INTERVENTIONS
03 — Hard Interventions

Hooks: Gates on the Pipeline

⚠️ Caveat: Hooks are only as good as their design. Overuse fills the context window and degrades soft instructions — the LLM starts spending tokens processing hook output rather than solving your problem. Use hooks for non-negotiable, high-risk actions only. Advisory rules belong in CLAUDE.md, not hooks.
SLIDE 04 / SOFT INTERVENTIONS
04 — Soft Interventions

Context = Life

88%
Healthy session: constraints stay resident and there is still room for the actual task.
Attention Is All You Need
CPU / RAM Analogy
CPU / attention24%
RAM / context28%
Bias Balanced
Action Keep working
Recency wins under pressure. The lower the water, the more bottom-of-file bias takes over.
Context Killers
SLIDE 05 / QUIZ
05 — Quick-Fire Quiz

Know Your Pitfalls?

QUESTION 1 OF 6
0 / 0
SLIDE 06 / BEST PRACTICES
06 — Claude Code Best Practices

Work Smarter
Not Just Faster

From Anthropic's own documentation — patterns that separate teams that ship from teams that drown in rework.

SLIDE 07 / VERDICT
07 — ROI Verdict

The Honest Verdict

Auto-coding is a force multiplier — but only with deliberate guardrails. The honest scorecard.

❌ Without Guardrails
🔴Context drift silently corrupts logic in long sessions
🔴Scope creep multiplies integration bugs undetected
🔴Brownfield tasks fail at 3× greenfield rate
🔴Supervision overhead wipes productivity gains
🔴Security surface expands without audit trail
ROI vs. CostNEGATIVE at 6mo
Delivery ConfidenceLOW
✅ With Hard + Soft Guardrails
🟢Context pinning keeps constraints consistent across turns
🟢Diff-scope hooks make scope creep physically impossible
🟢Test-first decomposition catches logic errors early
🟢Architecture rubric maintains pattern cohesion
🟢40–60% net velocity gain on greenfield tasks
ROI vs. Cost+40–60% greenfield
Delivery ConfidenceHIGH w/ oversight
Bottom line: Fast junior dev who writes a lot. Build the review pipeline before you build the feature.
Traceability Main site · pwdel.github.io #s0