Optimize

/optimize — Autonomous Optimization v2

What It Does

Runs an autonomous optimization loop against any target. The agent modifies the target, measures the result, keeps improvements, discards failures, and repeats until it stops climbing. Two modes: metric mode for code targets that produce a number (latency, bundle size), and eval mode for skills, prompts, or agents judged by LLM-as-judge binary evals.

The Problem

Tuning a thing for a measurable outcome is slow, boring, manual work. You change a file, run the measurement, eyeball whether it got better, keep or revert, then do it again — dozens of times. People give up after a few rounds and settle for "good enough" far short of the real ceiling. The targets without a clean number (a skill's quality, a prompt's effectiveness) are worse: there's no easy way to tell if a change actually helped. This skill runs that whole loop for you and only keeps changes that measurably win.

How It Works

Two modes drive the same hill-climb loop:

• Metric mode — code targets with a shell command that produces a number (the original).
• Eval mode — skills, prompts, agents, or any text target judged by LLM-as-judge binary evals.

Inspired by Karpathy's autoresearch and extended with LLM-as-judge evaluation.

Invocation

Metric Mode (code targets)

Copy code
/optimize --metric "lighthouse_score" --higher-is-better \
  --measure "npx lighthouse http://localhost:3000 --output=json" \
  --extract "jq '.categories.performance.score * 100' lighthouse.json" \
  --files "src/**/*.tsx,src/**/*.css" \
  --budget 120

/optimize --resume        # Resume a previous optimization loop
/optimize --status        # Show results summary from last/current run

Eval Mode (skill/prompt/agent targets)

Copy code
/optimize --target "~/.claude/skills/ExtractWisdom"
/optimize --target "~/.claude/skills/Research/Workflows/QuickResearch.md"
/optimize --target "prompts/my-prompt.md"
/optimize --target "~/.claude/skills/ExtractWisdom" --max-experiments 20

In eval mode, the system automatically:

1. Detects the target type (skill, prompt, agent, code, function)
2. Reads the target to understand its purpose and constraints
3. Generates 3-6 binary eval criteria and 3-5 test inputs
4. Presents criteria + inputs for your approval before starting
5. Runs the optimization loop using LLM-as-judge scoring
6. Presents a recommendation (apply/reject/partial) when done

What Happens

This skill drives the LifeOS Algorithm as an autonomous mutation loop:

1. OBSERVE — Define or auto-detect the target, set eval_mode
2. THINK — Analyze codebase/skill, generate hypothesis queue
3. PLAN — Prioritize hypotheses by expected impact
4. BUILD — Phase 0: TARGET ANALYSIS (see optimize-loop.md)
   • Detect target type, auto-generate eval criteria (eval mode), set up sandbox, baseline
5. EXECUTE — The autonomous loop (optimize-loop.md):
   • Hypothesize → Modify target → Measure (metric or eval) → Keep/Revert → Repeat
   • Metric mode: ~12 experiments/hour (at 5-min budget)
   • Eval mode: ~6-8 experiments/hour (multi-run judging is slower)
6. VERIFY — Phase 9: RECOMMEND — diff, summary, apply/reject/partial options
7. LEARN — Phase 10: EXTRACT LEARNINGS — what worked, what didn't, structured insights

Arguments — Metric Mode

Arguments — Eval Mode

Shared Arguments

Algorithm Integration

When /optimize is invoked, the eval_mode is set based on arguments (mode: is retired — never write it to frontmatter):

• --measure provided → eval_mode: metric (git branch sandbox)
• --target provided → eval_mode: eval (directory sandbox)

ISC criteria become guard rails — assertions that must hold true across ALL experiments. Guard rails must REMAIN satisfied perpetually. A violation triggers automatic revert regardless of score improvement.

Reference files:

• ~/.claude/LIFEOS/ALGORITHM/optimize-loop.md — the full loop protocol
• ~/.claude/LIFEOS/ALGORITHM/eval-guide.md — how to write good eval criteria
• ~/.claude/LIFEOS/ALGORITHM/archive/target-types.md — target detection and ISC generation

Examples

Metric Mode

Optimize page load time:

Copy code
/optimize --metric "lighthouse_perf" --higher-is-better \
  --measure "npx lighthouse http://localhost:3000 --output=json --output-path=lh.json" \
  --extract "jq '.categories.performance.score * 100' lh.json" \
  --files "src/**/*.tsx,src/**/*.css" \
  --target 95 --budget 120

Optimize bundle size:

Copy code
/optimize --metric "bundle_bytes" --lower-is-better \
  --measure "bun run build 2>&1 && du -sb dist/ | cut -f1" \
  --files "src/**/*.ts" \
  --constraints "all tests must pass"

ML training (Karpathy-style):

Copy code
/optimize --metric "val_bpb" --lower-is-better \
  --measure "uv run train.py > run.log 2>&1 && grep '^val_bpb:' run.log | cut -d' ' -f2" \
  --files "train.py" \
  --locked "prepare.py" \
  --budget 300

Eval Mode

Optimize a skill's Extract workflow:

Copy code
/optimize --target "~/.claude/skills/ExtractWisdom" --max-experiments 15

Optimize a standalone prompt:

Copy code
/optimize --target "prompts/summarize-article.md" --runs 5

Optimize with custom criteria:

Copy code
/optimize --target "~/.claude/skills/Research/Workflows/QuickResearch.md" \
  --criteria "Does the output contain specific facts with sources?" \
            "Is the output structured with clear sections?" \
            "Does the output avoid generic filler?" \
  --inputs "research quantum computing breakthroughs 2025" \
           "quick research on supply chain security" \
           "find recent developments in AI agents"

Gotchas

• Hill-climbing can get stuck in local optima. If score plateaus, consider resetting with different initial conditions.
• Eval mode vs metric mode: Use metric mode for quantifiable targets (latency, size). Use eval mode for qualitative targets (skill quality, prompt effectiveness).
• Regression tolerance prevents catastrophic changes. Don't set it to 0 — some regression in secondary metrics is acceptable if primary metric improves significantly.