breakthrough-loop

breakthrough-loop is a skill for Claude Code, Codex from GuoCheng24/breakthrough-harness. It costs 107 tokens per session (732 once invoked), scanned A, original, MIT.

A research workflow for testing many candidate methods against a fixed scoring system. It is aimed at projects trying to beat an existing baseline or state of the art.

In plain words
What is it for?
Building evaluation harnesses, generating and screening research ideas, comparing methods, and running progressively larger experiments.
Why use it?
It reduces wasted experiments by checking early whether the goal, data, evaluation method, and possible improvement are realistic.

Skill for Claude CodeCodex

Part of the breakthrough-harness plugin — 1 skill shipped together

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add skills/guocheng24/breakthrough-harness/breakthrough-loop
Any agent
npx skills add GuoCheng24/breakthrough-harness --skill breakthrough-loop
Clone the repo
git clone --depth 1 https://github.com/GuoCheng24/breakthrough-harness

Made for: Claude Code, Codex.

Or install breakthrough-harness, the plugin that ships this one along with the rest of its 1 skill.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for breakthrough-loop

README.md
[![agentmods](https://agentmods.dev/badge/skills/guocheng24/breakthrough-harness/breakthrough-loop.svg)](https://agentmods.dev/skills/guocheng24/breakthrough-harness/breakthrough-loop)
Your own site
<a href="https://agentmods.dev/skills/guocheng24/breakthrough-harness/breakthrough-loop"><img src="https://agentmods.dev/badge/skills/guocheng24/breakthrough-harness/breakthrough-loop.svg" alt="Measured on agentmods" height="20"></a>
Per session 107 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 732 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5.1 $0.00107 $0.00732
Opus 5 $0.00053 $0.00366
Sonnet 5 $0.00021 $0.00146
Haiku 4.5 $0.00011 $0.00073

Measured 5d ago against content hash 1ed7e158f151, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-05, from the pricing page.

Security

Grade A, and why

breakthrough-loop scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 5d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

.agents/skills/breakthrough-loop/SKILL.md · 63 lines

How it starts

The opening of the file, as written. The whole thing — 63 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Breakthrough Loop

Install: drop this folder into ~/.claude/skills/, or one command: /plugin marketplace add GuoCheng24/breakthrough-harness then /plugin install breakthrough-harness@breakthrough-harness. Full methodology: https://github.com/GuoCheng24/breakthrough-harness

Before anything: target triage (all four or stop)

  1. Cheap scoring function exists (seconds-minutes, unambiguous, machine-run)?
  2. Groundwork exists (data downloadable, baselines published)?
  3. A win is externally recognisable (leaderboard / referee-grade comparison)?
  4. Oracle ceiling + trivial/random baseline measured, gap worth mining? (Measure BEFORE round 1 - a flat ceiling discovered at round 10 is nine rounds of wasted budget.)

Phase 1 — build the harness first (not the method)

  • One evaluation entry point; ground truth stays inside the harness.
  • Tiered scoring: Tier0 seconds (shape/finiteness) → Tier1 minutes (small subset) → Tier2 hours (full set, survivors only).
  • Anti-cheat four: null models score first and must land on the floor; metric conventions pinned (double-report if the official one differs); calibration/evaluation data physically separated; an absurd baseline score freezes all conclusions.
  • Reproduce one published baseline number before trusting your own. If you cannot match it, you have not finished reading the recipe - keep reading (optimizer, loss, metric convention, operator - every layer moves numbers).
  • Verify every guard by breaking what it watches, and check it fails for the right reason.

Phase 2 — the loop (one round per day)

  1. Generate 10-30 variants, each with a one-line reason it might win; one-instance sanity check before entry.
  2. Sweep in parallel (subagents/processes as executors only - they call the entry point, they never write method code).
  3. Select on calibration data only.
  4. Confirm top-k on held-out data at full budget. Unreproduced gains do not exist. Persist numbers to disk, commit, and convert this round's failure modes into new harness checks.

Read the full file on GitHub · 63 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 5d ago First seen · 63 lines · 107 tokens per session scan A 1ed7e158f151

Subscribe to this mod's changes

breakthrough-loop is a skill published in the GitHub repository GuoCheng24/breakthrough-harness (1 stars, last pushed yesterday), licensed MIT. It adds 107 tokens to every session and 732 once invoked, about $0.0005 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.

Related

Other skills, from other repositories

planning-with-files

Persistent file-based planning for multi-step AI-agent work. Keeps taskplan.md, findings.md, and progress.md on disk; lifecycle hooks inject selected project planning context. Automatic recovery reads project planning files only. Explicit session-catchup.py --metadata reads same-project local agent session records and…

mxyhi/ok-skills · 117 tokens

infrastructure-publishing

Skill for the publishing infrastructure module providing academic publishing workflows including BibTeX CLI citation generation, APA/MLA citation helper functions, DOI management, Zenodo publication, arXiv submission preparation, GitHub releases, PyPI and TestPyPI package distribution, static-site deployment to GitHub…

docxology/template · 127 tokens

infrastructure-overview

Top-level skill for the research template infrastructure layer. Use in Cursor, Claude Code, or similar agents when editing or importing anything under infrastructure/, understanding the two-layer architecture, or wiring build/validation/rendering/publishing. Covers module discovery, import patterns, thin…

docxology/template · 80 tokens

infrastructure-validation

Skill for the validation infrastructure module providing PDF validation, markdown validation, output integrity checks, link verification, documentation audits, issue categorization, and repository scanning. Use when validating research outputs, checking document quality, running audits, or verifying cross-references.

docxology/template · 54 tokens

infrastructure-llm

Skill for the LLM infrastructure module providing local Large Language Model integration via Ollama. Covers client initialization, prompt templates, output validation, manuscript review generation, conversation context, and CLI usage. Use when querying LLMs, generating manuscript reviews, validating LLM outputs, or…

docxology/template · 66 tokens

research-workflow

Seven-stage research workflow (SCOPE→LITERATURE→REASON→DESIGN→COMPUTE→SYNTHESIZE→WRITE). Use for: structuring an AI agent's research process, generating literature review prompts, scoping methodology. Usage: from infrastructure.research import ResearchWorkflow; ResearchWorkflow.describe() Config: set stage overrides…

docxology/template · 91 tokens