Auto-Optimizer

Zero-friction optimization engine. Tell your OpenClaw agent "optimize my [thing]" and it runs autonomous improve-eval-keep/revert cycles until your metric goes up. Works on anything: copy, prompts, code, prediction strategies, configs.

Results in under 5 minutes. Zero configuration required for demos.

---

Quick Setup (AI Agent Instructions)

Follow these steps exactly. No judgment calls needed.

Step 1 — Install

CODEBLOCK0

Step 2 — Initialize a git repo (required)

The optimizer uses git to revert bad changes. Must be in a git repo.

CODEBLOCK1

Step 3 — Pick your mode

Scalar mode — you have a command that outputs a single number (test score, Brier score, conversion rate, word count, etc.)

Binary mode — you want to evaluate quality with yes/no criteria (copy quality, prompt effectiveness, UX clarity)

Step 4 — Run it

Scalar mode:
CODEBLOCK2

Binary mode:
CODEBLOCK3

Not sure? Use the wizard:

./skills/auto-optimizer/auto-optimizer.sh --wizard

---

Pre-Built Starter Packs

Three self-contained demos that run immediately. No files to create, no config needed.

Demo 1: Cold Outreach Optimizer

Optimizes a cold email template using a mock scoring metric (hook strength + clarity + CTA quality + length).

CODEBLOCK5

What it does:

• - Creates a sample outreach template in INLINECODE0
• Runs a mock metric that scores: hook ≤15 words, single CTA, body ≤120 words, value prop clarity
• Iterates 5 times, keeps improvements, reverts regressions
• Prints a final report showing baseline → best score

Sample outreach template used:
CODEBLOCK6

Mock metric logic (inline in demo):
CODEBLOCK7

Demo 2: Prediction Market Strategy

Runs the optimization loop on a prediction strategy file, scoring by mock accuracy.

CODEBLOCK8

What it does:

• - Creates a sample prediction strategy in INLINECODE1
• Runs a mock metric that scores: specificity of criteria, use of base rates, calibration language
• Shows how the loop works on structured reasoning files

Demo 3: Prompt Quality Optimizer (Binary Mode)

Optimizes a system prompt using 5 yes/no quality criteria.

CODEBLOCK9

What it does:

• - Creates a sample system prompt in INLINECODE2
• Evaluates each iteration against 5 criteria (inline):

1. Does the prompt specify a clear role/persona?
2. Does it include explicit output format instructions?
3. Does it define what NOT to do?
4. Is it under 500 words?
5. Does it include at least one concrete example?

• - Batch size 10: generates 10 outputs, scores each, calculates pass rate %
• Keeps versions that increase the pass rate

Sample system prompt used:

You are a helpful assistant. Answer questions clearly and accurately.
Be concise but thorough. Help the user accomplish their goals.

---

Full Capability Guide

--wizard — Interactive Setup

Walks you through setup interactively. Best when you're not sure which mode to use.

CODEBLOCK11

Prompts you to choose:

1. 1. Cold outreach / email copy → sets up binary mode with outreach evals
2. LLM prompt / system prompt → sets up binary mode with prompt quality evals
3. Prediction market strategy → sets up scalar mode with accuracy metric
4. Code / config file → sets up scalar mode, prompts for your test command
5. Custom → asks for your file and metric

--eval-mode scalar (default)

When to use: Anything with a measurable number. Test pass rate, Brier score, word count, latency, revenue, API response score.

Requirements: Your --metric command must print a single float to stdout.

CODEBLOCK12

How it works: Runs metric → agent proposes change → run metric again → if improved, commit; else git checkout to revert.

--eval-mode binary

When to use: Copy, prompts, UX, anything where quality is multi-dimensional and hard to reduce to one number.

Requirements: An evals file (markdown list of yes/no criteria) and a --batch-size (default 10).

CODEBLOCK13

How it works: For each iteration, generates batch-size outputs from the current file, scores each against all criteria, calculates overall pass % → agent proposes change → compare pass % → keep or revert.

--budget N

Number of optimization iterations to run. Each iteration = one agent call + one eval.

Budget	Time (approx)	Best for
5	~2 min	Quick demo, sanity check
10

~5 min | Initial optimization pass |
| 20 | ~10 min | Production runs |
| 50+ | ~30 min | Overnight deep optimization |

Minimum effective budget: 5 iterations. Below 5, not enough signal.

--goal minimize / --goal maximize (default: maximize)

CODEBLOCK14

--session NAME

Name your session for organized results. Results saved to ./skills/auto-optimizer/results/NAME/.

CODEBLOCK15

--batch-size N (binary mode only)

How many outputs to generate per iteration for scoring. Higher = more reliable signal, slower.

• - 5 = fast, less reliable
• 10 = balanced (default)
• 20 = slow, high confidence

Hypothesis Memory

Every iteration is logged to results/SESSION/hypothesis_log.jsonl. The agent reads the last 5 entries before each iteration, so it never retries approaches that already failed.

This is what makes multi-iteration runs productive rather than random. The optimizer builds on what worked, avoids what didn't.

---

OpenClaw Integration

The natural way (just tell your agent)

CODEBLOCK16

Your OpenClaw agent reads this SKILL.md, picks the right mode, sets up the files, and runs the loop.

Direct invocation

CODEBLOCK17

---

Real Results (What to Expect)

From actual runs:

Prediction market strategy — 5 iterations:

• - Brier score: 0.23642 → 0.23563 (↓ better)
• ROI: +2.83% → +6.25%
• What changed: added base rate anchoring, tightened confidence thresholds

Cold outreach template — 10 iterations:

• - Binary pass rate: 60% → 85%
• Version: v1.0 → v1.1
• What changed: shorter hook (22 words → 11 words), single clear CTA, ICP-specific pain point added

System prompt — 20 iterations (binary):

• - Pass rate: 40% → 92%
• What changed: added explicit persona, output format section, negative constraints, inline example

Typical pattern:

• - Iterations 1-3: big gains (low-hanging fruit)
• Iterations 4-10: diminishing returns, more targeted changes
• Iterations 10+: fine-tuning, marginal improvements

---

Troubleshooting

"Not a git repo"
CODEBLOCK18

"Metric command failed" or returns 0 always
Your metric command must print a single float to stdout. Test it standalone:

bash ./your-metric.sh
# Should output: 73.5

If it outputs anything else (multiline, text, nothing), wrap it:
CODEBLOCK20

"claude CLI not found"
Option A: Install claude CLI globally: npm install -g @anthropic-ai/claude-code
Option B: The script falls back to OpenClaw's claude-code skill automatically if skills/claude-code/claude-code.sh exists.

"ERROR: --evals is required for binary eval mode"
Binary mode needs an evals file with numbered criteria:
CODEBLOCK21

"Budget too low"
Minimum 5 iterations to see meaningful improvement. Use --budget 10 for first runs.

Results not improving after many iterations

• - Check hypothesis log: INLINECODE20
• The agent may be stuck in a local optimum — try a new session with INLINECODE21
• Consider rewriting the evals/metric to be more discriminating

"program.md.template not found"

ls skills/auto-optimizer/
# Should show: auto-optimizer.sh, SKILL.md, program.md.template, results/

If missing, reinstall: clawhub install auto-optimizer

---

MiroFish Integration

MiroFish is a swarm intelligence engine that runs thousands of AI agents to simulate outcomes and generate prediction reports. Combined with auto-optimizer, you can autonomously improve the "seed" inputs that drive MiroFish simulations.

What MiroFish does

• - Takes a "seed" (market data, news, signals) as input
• Runs multi-agent social simulation in a digital world (Twitter + Reddit environments)
• Simulates distinct personas: RetailTrader, WhaleInvestor, AlgorithmicTrader, etc.
• Outputs agent actions, discussions, and quantitative probability estimates

The combination

• - Mutable asset: the seed file / prediction prompt sent to MiroFish
• Scalar metric: confidence score OR prediction accuracy on known outcomes
• Loop: auto-optimizer iterates on the seed to maximize simulation confidence

Proven optimization results (2026-03-26 run)

Iteration	Seed Content	Confidence Score	Delta
Baseline	Price + Fear/Greed only	0.35	—
Iter 1

+ Technical signals (TTM Squeeze, funding rates) | 0.58 | +0.23 ✅ |

| Iter 2 | + Whale on-chain data + price levels | 0.67 | +0.09 ✅ | | Iter 3 | + Cross-asset correlation (BTC) | 0.61 | specific question |

Key insight: Adding technical structure data (funding rates, squeeze, key levels) produces the biggest confidence boost. Whale on-chain context is the #2 improvement driver.

Setup

MiroFish must be running (backend on port 5001): CODEBLOCK23

Build a seed from live market data:
CODEBLOCK24

Run a simulation manually via the API:
CODEBLOCK25

Scoring function for auto-optimizer

Extract confidence from MiroFish agent actions: CODEBLOCK26

Use cases

• - Crypto market predictions — BTC/ETH/SOL price direction (24-72h)
• Prediction market research — Polymarket/Kalshi question research
• Seed quality optimization — find which data signals drive the highest swarm confidence
• Any "what if" scenario you want to simulate at scale with diverse AI personas

Files

• - Seed builder: INLINECODE23
• Results example: INLINECODE24
• MiroFish API: http://localhost:5001 (Flask backend, port 5001)