Stage 1/frame-optimization
frame‑optimization
Defines the objective, measurement rules and allowed search space, then obtains the required independent review before any campaign starts.
PROBLEM.mdREPRESENTATION.md
NoeHub Factory Optimization · Module 01Aka Optimization
Push the frontier
Aka Optimization is an optimization research workflow for coding agents. Define what better means, test promising ideas, and let the evidence choose the next improvement.
When tests aren’t enough
Passing tests can’t tell you whether a change is better.
Aka Optimization gives your coding agent a research discipline. Goals, evaluation rules, budgets and decisions live in your repository, so results stay comparable and another session can pick up exactly where the last one stopped.
- Inference cost
- Latency
- Model quality
- Game-playing agents
- Configuration tuning
- Alternative implementations
How it works
One loop.
Evidence decides
the next move.
The workflow aims for the best attainable result. It encourages ambitious hypotheses, compares alternative approaches, and chooses the test that resolves the uncertainty that matters most to the next decision.
-
01
Define the goal and comparison rules
What are we improving, what may change, and how is a result measured and compared?
-
02
Decide what can be changed and searched
Turn the allowed solution space into a representation a search can actually move through.
-
03
Research, implement, and measure
Pick the highest-leverage affordable test — the one that can most quickly falsify a hypothesis or expose a better route.
-
04
Improve, change direction, or stop
A local optimum is provisional. Results feed the next cycle — including the search strategy itself.
Two coordinators
Frame the problem.
Then run the campaign.
You normally invoke only two Skills. They assign specialists for research, design, implementation, review — and the questions that genuinely need your judgment.
Stage 2/frontier-optimization
frontier‑optimization
Runs the improvement campaign within the agreed limits, learns from every result, and preserves the context needed to resume.
FRONTIER.mdfrontier/
RSI This is Recursive Self-Improvement: evidence from one cycle informs the next. Where the task permits it, that includes improving the search strategy, the evaluator, or the tools — while changes to comparison rules stay explicit, so incompatible results are never compared.
Thirteen Skills, one workflow
Specialists that
never overrule the contract.
Workers research, design, ask, review and reflect — each in a bounded assignment. None of them can silently change the task definition or the search design. Install all thirteen together.
A Define the task and search
frame-optimizationCoordinates the problem definition, search representation, repair loops, reviews and handoff.design-measurementDesigns reusable measurement meaning in a fresh context.research-optimizationInvestigates one assigned question without changing the contract.grill-optimizationCollects one decision that only you can make.review-optimizationIndependently checks problem and measurement readiness.review-representationIndependently checks the search space later work may use.
B Run the improvement campaign
frontier-optimizationCoordinates entry, planning, budget, batch selection, adoption, closeout and recovery.design-implementationResolves a consequential technical design question.research-frontierSource-grounded, multidisciplinary research — never a closed menu.grill-frontierAsks one tradeoff or permission question, only when evidence can’t decide.run-frontier-batchImplements, debugs, integrates and observes assigned Batch work.review-frontierReviews exact Git-backed designs, implementations and claims.reflect-frontierTurns route experience into insight and promising next ideas.
What to expect
Built for campaigns
that outlive a session.
Comparable results
The task states what may change, how it is measured, and which conclusions the evidence supports. Results carry the problem and representation versions they were produced under.
Learning across experiments
Failed ideas, useful alternatives and reasons for changing direction remain available to later work — a failed bold hypothesis still eliminates uncertainty.
Work within your limits
The agent reuses decisions you already made and asks only when a paid, external, sensitive or irreversible action needs authorization you haven’t given.
Recoverable progress
Project files retain the current question, the next observation, resource use, and conditions for continuing or stopping. State is recovered from Git, not chat history.
Git
Evidence lives in your repo. Reviews name exact commits and paths; restoring files never undoes spending or effects already incurred.
Rust
Optional execution harness. Preserves bounded assignments across agent calls and recovers without repeating uncertain work.
Hook
Optional Codex Hook. Restores saved work context and observes agent dispatch. It grants no permission and never auto-continues.
Honest limit This is a research workflow, not a domain-specific optimizer. Installing it and passing its checks does not prove it will improve your system — that takes evidence from your task.
Get started
Three steps to your
first campaign.
Install the Skills
Install all thirteen together, in a Git repository with Python 3.10+ and PyYAML 6.x. For the Codex desktop app, add the plugin and restart Codex:
codex plugin marketplace add Noehub2025/noehub-factory-optimization
codex plugin add noehub-factory-optimization@noehubFor Claude Code, run inside a session:
/plugin marketplace add Noehub2025/noehub-factory-optimization
/plugin install noehub-factory-optimization@noehubFor other Agent Skills–compatible agents:
npx skills@latest add Noehub2025/noehub-factory-optimizationAdd the project instructions
In Codex, the plugin's setup Skill does this for you: ask it to set up your project. For other agents, copy AGENTS.md into your project root, merging with existing rules. Claude Code imports it through one line in CLAUDE.md:
@AGENTS.mdStart a fresh agent session and confirm the instructions are loaded.
Start a task
Describe your objective and ask for both stages:
I want to reduce inference cost while preserving the quality we agree to measure. Use frame-optimization to define and review inference-cost-reduction, then continue with frontier-optimization within the adopted objective, budget, access, and effect limits.The task is saved under docs/skills/optimization/<task>/ so later sessions can resume it. It plans and measures before changing code, so even a small task takes 15 to 30 minutes and may create Git commits.
NoeHub Factory Optimization
Modules that make
teams more capable.
Factory Optimization is NoeHub’s series of modules for raising what a company’s teams can get done. Each module is one focused workflow that works inside the tools your team already uses. Aka Optimization is the first.
01
Available · v1.2.1
02
In development
Aka Optimization
Optimization research for coding agents. Define what better means, test promising ideas, and let the evidence choose the next improvement.
Get startedNext module
More Factory Optimization modules are on the way.
About NoeHub
From insight to impact —
quietly crafted software.
NoeHub is a small studio in Hangzhou making focused tools for careful work. Its Factory Optimization modules bring that care to how companies get work done.
- Legal entity
- Hangzhou Boshu Technology Co., Ltd.
- Type
- Limited Liability Company
- Based in
- Hangzhou, Zhejiang, China
- Founded
- 14 May 2025
- Privacy
- Privacy notice
+ Coming from NoeHub
Career Development Report
A structured assessment for knowledge workers in their first ten years — one priority direction, with the evidence behind it.