Agent skill

Vs Meta Opt

by uw-syfi in uw-syfi/vibesys

Run and audit long-lived VibeSys meta-optimization campaigns that improve VibeSys as an optimizer rather than directly optimizing its bespoke candidate.

MITAuto-check passedBusiness, Finance & HR

Install Vs Meta Opt

skills CLI
$ npx skills add uw-syfi/vibesys --skill vs-meta-opt -a claude-code

Project install by default; add -g for ~/.claude/skills/.

GitHub CLI
$ gh skill install uw-syfi/vibesys vs-meta-opt --agent claude-code

Project scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).

Manual copy
$ git clone --depth 1 https://github.com/uw-syfi/vibesys.git skills-src && mkdir -p .claude/skills && cp -r skills-src/.agents/skills/vs-meta-opt .claude/skills/vs-meta-opt && rm -rf skills-src

Use ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.

Claude Code skills documentation · loads skills from .claude/skills/

Facts

Skill name
vs-meta-opt
GitHub stars
108
Token cost
~2.9k tokens
SKILL.md length
1,484 words
Files
8 (incl. references, assets)
Skills in repo
15
Repo updated
First seen
Licence
MIT

At a glance

Run and audit long-lived VibeSys meta-optimization campaigns that improve VibeSys as an optimizer rather than directly optimizing its bespoke candidate.

  • Works in 8 steps: Open the draft PR → Launch or resume VibeSys → Monitor with adaptive heartbeats → …
  • Codex should launch
  • SKILL.md covers Purpose, Read Only What Is Relevant, Run Contract and Control Loop, plus 3 more sections
  • Instructions only: no scripts, shell commands, URLs or credentials in SKILL.md

What it does

Vs Meta Opt is an agent skill from uw-syfi/vibesys. Run and audit long-lived VibeSys meta-optimization campaigns that improve VibeSys as an optimizer rather than directly optimizing its bespoke candidate. Use when Codex should launch or resume VibeSys with a specified input, monitor every completed round, use independent counterfactual reviews when a trajectory plateaus or becomes uncertain, diagnose ineffective agent-system behavior, implement and commit VibeSys changes, compare later trajectory windows, and repeat within round, time, cost, or intervention budgets.

Its SKILL.md is about 2.9k tokens, which your agent loads only when the skill is triggered. The skill folder holds 10 other files, including reference files and assets (for example `agents/openai.yaml`, `assets/meta-run-pr-template.md` and `references/counterfactual-review.md`).

It sits in Business, Finance & HR. The repository describes itself as: Can AI Agents Build Bespoke Systems? The licence is MIT.

When your agent uses it

  • Codex should launch
  • Resume VibeSys with a specified input
  • Monitor every completed round
  • Use independent counterfactual reviews when a trajectory plateaus

Example prompts

  • “/vs-meta-opt”

Workflow steps

8 steps, taken from the step headings in SKILL.md.

  1. Open the draft PR
  2. Launch or resume VibeSys
  3. Monitor with adaptive heartbeats
  4. Check the trajectory after every round
  5. Audit and form the next meta-hypothesis
  6. Pause, change, validate, and commit
  7. Resume and measure the effect
  8. Repeat within budget

What it can do on your machine

Read from SKILL.md and the folder at commit 999938a. It shows what the files ask for, not the result of running them.

  • Tool permissions

    Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.

    From allowed-tools in the SKILL.md frontmatter.

  • Runs code

    No scripts in the folder and no shell commands in SKILL.md.

    From the folder's file list and the shell code blocks in SKILL.md.

  • Network

    No URLs in SKILL.md.

    From URLs in SKILL.md, links to its own repository left out.

  • Credentials

    Names no API keys, tokens, secrets or passwords.

    From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.

Context cost

Vs Meta Opt loads about 2.9k tokens when it runs, and up to ~11k if it reads all its reference files. Until then it costs about 133 tokens; SKILL.md has 1,484 words of instructions outside code blocks.

Always · name and description, kept in context so the agent knows when to use it
~133
When it runs · the whole SKILL.md, loaded when a task matches
~2.9k
With references · SKILL.md plus every file in references/, read only if the agent opens them
~11k

Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.

Safety

Auto-check passed

The automated check found no risky patterns in SKILL.md.

Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); files beside SKILL.md are not scanned.

SKILL.md

The full file from uw-syfi/vibesys at commit 999938a, republished under its MIT licence (© uw-syfi). 1,484 words, ~2,925 tokens.

Download SKILL.mdSave it as .claude/skills/vs-meta-opt/SKILL.md (or your agent's skills folder). This skill also uses 7 other files; get the full folder from GitHub.
name
vs-meta-opt
description
Run and audit long-lived VibeSys meta-optimization campaigns that improve VibeSys as an optimizer rather than directly optimizing its bespoke candidate. Use when Codex should launch or resume VibeSys with a specified input, monitor every completed round, use independent counterfactual reviews when a trajectory plateaus or becomes uncertain, diagnose ineffective agent-system behavior, implement and commit VibeSys changes, compare later trajectory windows, and repeat within round, time, cost, or intervention budgets.

VS Meta Opt

Purpose

Operate a feedback loop that judges and improves the optimizer, not the bespoke system it produces. Run VibeSys, observe whether its search is effective, change VibeSys when evidence supports a meta-hypothesis, resume the campaign under the new version, and determine whether the trajectory improved.

Preserve a strict boundary:

  • Treat candidate implementation choices and domain performance techniques as inner-loop work.
  • Treat role behavior, search policy, prompts, skill availability, evidence, evaluation, context, lifecycle, and cost control as meta-optimization surfaces.
  • Do not become a shadow candidate designer.

Read Only What Is Relevant

Read campaign state, history, prompts, diffs, logs, profiles, timings, costs, and evaluations from artifact paths. Read deltas after the initial inspection; do not repeatedly inject durable history or unchanged output into context.

Run Contract

Before starting, record in the draft PR:

  • Exact VibeSys command, specified input paths and revisions, environment, and initial VibeSys commit.
  • Campaign identity, initial checkpoint, and correctness/evaluation contract.
  • Maximum rounds, wall-clock time, accelerator time or cost, agent budget when applicable, and meta-interventions.
  • A terminal reserve for final official evaluation, evidence publication, and cleanup.
  • Initial trusted frontier and available effectiveness measurements.

Treat one VS Meta Opt run as one bounded meta-optimization campaign. It owns one branch and one draft PR, but may contain multiple VibeSys process segments, sequential meta-hypotheses, clean commits, validations, and explicit reverts.

Control Loop

1. Open the draft PR

Start from the intended VibeSys parent on a dedicated clean branch. Open the draft PR before changing behavior and initialize meta-run-pr-template.md. Use its description as the living effectiveness ledger.

2. Launch or resume VibeSys

Run VibeSys with the specified input and exact recorded command. Preserve the campaign checkpoint and candidate state across process restarts. Record the VibeSys commit governing each campaign segment.

Use the simple edit path by default: finish the current round, stop or pause VibeSys, edit and commit in the same checkout, then resume. When overlapping work is worthwhile, optionally develop in a separate clean worktree while leaving the active framework checkout unchanged. Treat every framework change as inactive until a recorded campaign segment starts or resumes under its commit.

3. Monitor with adaptive heartbeats

Choose the next sleep interval from the current phase, expected duration, progress signal, failure risk, and budget. Check sooner during startup, near a deadline, or after a warning; sleep longer during healthy work that is expected to take time. Avoid busy polling.

After each sleep, inspect only new status, events, summaries, or a bounded log tail. If nothing changed and the process is healthy, do not rerun semantic analysis or reinsert unchanged output.

Perform at least one trajectory check after every completed VibeSys round. Between round completions, use heartbeats only to verify liveness, phase progress, deadlines, and resource safety.

4. Check the trajectory after every round

Use a lightweight per-round check to answer:

  • Did the round produce valid new evidence or frontier progress?
  • Did the next decision respond to the evidence available?
  • Is VibeSys repeating a known failure, tuning within noise, or losing context?
  • Did evaluation, coordination, or infrastructure consume disproportionate work?
  • Are correctness, provenance, and reward-hacking defenses intact?
  • Does enough budget remain for the current path or another meta-experiment?

If the trajectory remains healthy, record a compact checkpoint and continue without proposing a change. Trigger a deeper audit when a warning recurs, the trajectory plateaus, integrity is uncertain, cost rises unexpectedly, or the current validation window ends.

When the next-step quality is uncertain—especially during a plateau—consider a counterfactual trajectory review using counterfactual-review.md. Give fresh subagents the legitimate objective and raw artifact paths, but withhold VibeSys's proposed next step until they produce independent alternatives. Do not run this panel automatically after every round.

5. Audit and form the next meta-hypothesis

Apply effectiveness-rubric.md at round, window, and campaign timescales. Classify findings, including Other when nothing fits cleanly, and identify the narrowest VibeSys-owned mechanism supported by evidence.

When a counterfactual review exists, compare its proposals with VibeSys's choice using evidence fit, causal model, expected impact, falsifiability, cost, integrity, and novelty. Do not assume the subagents are right. If a stronger alternative exposes a proposal gap, explain why VibeSys missed it using the failure taxonomy before changing anything.

Propose one next system intervention with expected meta-metric effects, likely regressions, a validation window, success criteria, and a reversion condition. Improve how VibeSys generates or evaluates directions; do not paste the stronger candidate proposal into an always-on prompt. If evidence is insufficient, improve instrumentation before changing behavior.

Show full SKILL.md (671 more words)Show less
6. Pause, change, validate, and commit

Stop or pause at a durable round boundary. Keep candidate code and campaign artifacts separate from the VibeSys change. Implement the narrowest coherent design that fully addresses the meta-hypothesis; scope by causal responsibility, not line count.

Follow repository conventions and architecture boundaries. Reuse or extend existing abstractions when they own the behavior, and introduce or refactor abstractions when that materially improves ownership, data flow, reuse, testability, or ergonomics. Include adjacent cleanup required for a natural design, but exclude unrelated refactoring. Reject one-off conditionals, compatibility shims, and other patchwork used only to minimize the diff.

Run targeted checks, perform the leakage review, and make a clean rationale-bearing commit.

When the intervention changes a schema for VibeSys-generated code or persisted run state, snapshot the existing state, update the canonical schema and all producers and consumers, migrate the live campaign once, validate the result, and resume using only the new representation. Commit the schema, one-way migration, and tests atomically. Do not retain parallel runtime schema implementations.

Keep prompts procedural and neutral. Do not encode a candidate optimization, known bottleneck, benchmark-specific trick, prior winning implementation, or hidden evaluator behavior directly into prompts. Adding or improving modular, versioned, selectively loaded skills is allowed and must be recorded as a capability change.

Update the PR intervention log and remaining budget immediately.

7. Resume and measure the effect

Resume the same campaign from the recorded checkpoint under the new VibeSys commit. Record the new segment boundary. Observe the predeclared number of rounds or other validation condition before attributing an effect, unless correctness, safety, or decisive contrary evidence requires stopping early.

Use framework-owned evidence to retain, revise, or explicitly revert the intervention. Do not certify it merely because its author expected it to work.

8. Repeat within budget

Continue monitoring, per-round trajectory checks, meta-hypotheses, commits, and validation until the budget reaches its terminal reserve or another stopping condition fires. Do not begin an intervention without enough remaining budget to evaluate it.

Finish with the terminal official evaluation when enabled, publish final evidence, terminate owned processes and remote resources, update the PR disposition, and leave the worktree clean.

Token-Efficient Monitoring Rules

  • Let the agent choose sleep intervals contextually; do not impose one global cadence.
  • Never skip the trajectory check after a completed round.
  • Track the last event or log cursor, round, phase, progress time, VibeSys commit, remaining budget, and next wake reason.
  • Read append-only deltas and compact summaries before raw logs.
  • Keep complete logs on disk and load only evidence needed for the current decision.
  • Separate cheap heartbeat/liveness checks from expensive semantic audits.
  • Use counterfactual subagents only when their expected information value justifies their token and time budget.
  • Treat silence according to phase-specific expectations; do not call healthy long work a hang.
  • Capture bounded diagnostics before terminating a stalled process, then clean up every owned resource.

Guardrails

  • Do not propose first-order candidate optimizations as the audit result.
  • Do not put discovered optimizations or tricks into always-on agent prompts.
  • Do not judge VibeSys only by the final candidate score.
  • Do not force every finding into a closed taxonomy; use Other with rationale.
  • Do not infer causality from one successful round without considering opportunity, cost, and confounders.
  • Do not mix a VibeSys system intervention with candidate changes in one commit.
  • Do not optimize for the fewest changed lines at the expense of an ergonomic, project-conventional design.
  • Do not mutate the active framework checkout; stop first or use an optional separate worktree. Generated code and state may be migrated at a safe boundary to one new canonical schema.
  • Do not spend the terminal reserve on a change that cannot be evaluated.

Expected Handoff

Return:

  1. Run command, input, campaign checkpoint, active VibeSys commit, and evidence paths.
  2. Per-round and window effectiveness findings, including Other when applicable.
  3. Current phase, last progress, next wake reason, and remaining budgets.
  4. Limiting VibeSys mechanism and the next intervention with leakage assessment.
  5. Commit and campaign-segment log with validation dispositions.
  6. Final effectiveness assessment, terminal evaluation, cleanup status, and unresolved confounders.

© uw-syfi, MIT. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file

Files

SKILL.md and 7 other files (references, assets) in .agents/skills/vs-meta-opt of uw-syfi/vibesys.

  • SKILL.md
  • agents/openai.yaml
  • assets/meta-run-pr-template.md
  • references/counterfactual-review.md
  • references/diagnosis-and-levers.md
  • references/effectiveness-rubric.md
  • references/meta-experiments-and-commits.md
  • references/run-control.md

Open the folder on GitHubat commit 999938a

Compare with similar skills

Vs Meta Opt next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.

Vs Meta Opt compared with similar skills
SkillStarsUsed inTokensAuto-checkLicenceRepo updated
Vs Meta Opt this skilluw-syfi/vibesys108—~2.9kAutomated safety check: PassMIT
Technical Analysttradermonty/claude-trading-skills3k4 repos~4.6kAutomated safety check: PassMIT
Theme Detectortradermonty/claude-trading-skills3k2 repos~4.9kAutomated safety check: PassMIT
Creating Financial ModelsChen-zexi/open-ptc-agent7293 repos~1.3kAutomated safety check: PassMIT
Stock APIzhangxiangliang/stock-api2k—~507Automated safety check: PassMIT
Itr Walakaranb192/itr-wala871—~3.6kAutomated safety check: PassMIT

Similar skills

  • Technical Analyst

    tradermonty/claude-trading-skills

    This skill should be used when analyzing weekly price charts for stocks, stock indices, cryptocurrencies, or forex pairs.

    3k GitHub starsUsed in 4 repos~4.6k tokens
    Business, Finance & HRAuto-check passed
  • Theme Detector

    tradermonty/claude-trading-skills

    Detect and analyze trending market themes across sectors. An agent skill from tradermonty/claude-trading-skills.

    3k GitHub starsUsed in 2 repos~4.9k tokens
    Business, Finance & HRAuto-check passed
  • Creating Financial Models

    Chen-zexi/open-ptc-agent

    This skill provides an advanced financial modeling suite with DCF analysis, sensitivity testing, Monte Carlo simulations, and scenario planning for investment decisions

    729 GitHub starsUsed in 3 repos~1.3k tokens
    Business, Finance & HRAuto-check passed
  • Stock API

    zhangxiangliang/stock-api

    Fetch real-time stock quotes, K-line (candlestick) history, and search symbols for China A-shares, Hong Kong, and US markets.

    2k GitHub stars~507 tokensUpdated yesterday
    Business, Finance & HRAuto-check passed
  • Itr Wala

    karanb192/itr-wala

    File Indian income tax returns (ITR) for FY 2025-26 / AY 2026-27.

    871 GitHub stars~3.6k tokensUpdated 7 days ago
    Business, Finance & HRAuto-check passed
  • Tushare Data

    zillionare/zillionare

    面向中文自然语言的 Tushare 数据研究技能。用于把“看看这只股票最近怎么样”“帮我查财报趋势”“最近哪个板块最强”“北向资金在买什么”“给我导出一份行情数据”这类请求,转成可执行的数据获取、清洗、对比、筛选、导出与简要分析流程。适用于 A 股、指数、ETF/基金、财务、估值、资金流、公告新闻、板块概念与宏观数据等研究场景。

    322 GitHub starsUsed in 2 repos~2.3k tokens
    Business, Finance & HRAuto-check passed

More from uw-syfi/vibesys

All 15 skills in this repo
  • Neuron Nki Profiling

    uw-syfi/vibesys

    This skill guides using the cli to generate NKI kernel profiles (NEFF + NTFF pairs) to analyze performance on Neuron hardware.

    108 GitHub stars~3.2k tokensUpdated today
    Auto-check passed
  • Neuron Nki Debugging

    uw-syfi/vibesys

    This skill guides debugging NKI compilation errors on Neuron hardware.

    108 GitHub stars~2.7k tokensUpdated today
    Auto-check passed
  • Neuron Nki Docs

    uw-syfi/vibesys

    Research NKI documentation for API lookups, tutorials, error codes, architecture, and optimization guides.

    108 GitHub stars~1.9k tokensUpdated today
    Auto-check passed
  • Query and analyze NKI kernel profile data from neuron-explorer parquet files.

    108 GitHub stars~4.3k tokensUpdated today
    Auto-check passed
  • Neuron Nki Writing

    uw-syfi/vibesys

    Guide for writing and modifying NKI kernels. An agent skill from uw-syfi/vibesys.

    108 GitHub stars~5k tokensUpdated today
    Auto-check passed
  • Triage PRs

    uw-syfi/vibesys

    Triage the open pull requests of the VibeSys repository. An agent skill from uw-syfi/vibesys.

    108 GitHub stars~1.2k tokensUpdated today
    Auto-check passed

Questions about Vs Meta Opt

What does Vs Meta Opt do?

Run and audit long-lived VibeSys meta-optimization campaigns that improve VibeSys as an optimizer rather than directly optimizing its bespoke candidate. Vs Meta Opt is an agent skill from uw-syfi/vibesys. Run and audit long-lived VibeSys meta-optimization campaigns that improve VibeSys as an optimizer rather than directly optimizing its bespoke candidate.

When should I use Vs Meta Opt?

Vs Meta Opt fits situations like: Codex should launch; resume VibeSys with a specified input; monitor every completed round; use independent counterfactual reviews when a trajectory plateaus.

How do I install Vs Meta Opt in Claude Code?

Run `npx skills add uw-syfi/vibesys --skill vs-meta-opt -a claude-code`. Or copy the skill folder (.agents/skills/vs-meta-opt in uw-syfi/vibesys) into .claude/skills/vs-meta-opt in your project. Claude Code loads it when a task matches its description.

How do I install Vs Meta Opt in Codex?

Run `npx skills add uw-syfi/vibesys --skill vs-meta-opt -a codex`. Or copy the skill folder (.agents/skills/vs-meta-opt in uw-syfi/vibesys) into .agents/skills/vs-meta-opt in your project. Codex loads it when a task matches its description.

Can I use Vs Meta Opt in Cursor, Gemini CLI or GitHub Copilot?

Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add uw-syfi/vibesys --skill vs-meta-opt -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/vs-meta-opt, .gemini/skills/vs-meta-opt, .github/skills/vs-meta-opt and .opencode/skills/vs-meta-opt in your project.

What does Vs Meta Opt need to run?

SKILL.md names no scripts, command-line tools or credentials: Vs Meta Opt is instructions for the agent only.

Does Vs Meta Opt access the network?

SKILL.md contains no URLs. Any network use would come from the scripts or tools the agent runs. This is read from the text; nothing was executed.

Is Vs Meta Opt safe to install?

Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. Review the folder before installing.

What licence does Vs Meta Opt use?

Vs Meta Opt is published under the MIT licence (the repository's licence). It allows redistribution, so the full SKILL.md is shown on this page.

How many tokens does Vs Meta Opt use?

About 2.9k tokens (SKILL.md is roughly 12k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full. Its references folder adds about 8.3k tokens, read only when the agent opens those files.

What are the alternatives to Vs Meta Opt?

Skills that share tags, products or a category with Vs Meta Opt: Technical Analyst (tradermonty/claude-trading-skills, 3k stars), Theme Detector (tradermonty/claude-trading-skills, 3k stars), Creating Financial Models (Chen-zexi/open-ptc-agent, 729 stars) and Stock API (zhangxiangliang/stock-api, 2k stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.

Who maintains Vs Meta Opt?

uw-syfi (a GitHub organization) maintains it in uw-syfi/vibesys, which has 108 GitHub stars. The repository holds 15 skills in this directory. The repository was last updated on October 11, 2026.

Source: uw-syfi/vibesys on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.