maxforge-lab

Maxforge Lab

AI tooling, forged for business. Maxforge Lab builds and hardens production-grade agent capabilities — from packaged skills to bespoke, high-end AI solutions for teams that need results they can trust.

This repository is the Lab's flagship open toolkit: the research-to-execution skill set for coding agents. Four composable agent skills that take an open-ended goal from not knowing → knowing → a written plan → verified action.

Built for Claude Code (and other skill-aware harnesses). Each skill is a behavior-shaping reference the agent loads on demand — and every one ships with a documented eval so you know it works before you bet a workflow on it.

Working with Maxforge Lab. The skills here are free and MIT-licensed. If you need custom skills, agent workflows tuned to your domain, or a hardened deployment on your own model stack, that's the Lab's high-end engagement tier — reach out at [email protected].

---

The four skills

Skill	Use it when…	What it enforces
researching-topics	You need to gather, evaluate, and synthesize information before acting	Frame the question → gather from independent sources (in parallel) → corroborate → synthesize → cite → state confidence + the key assumption you're betting on (even under "no hedging" pressure)
analyzing-data	You have data of any kind and need insight (descriptive, diagnostic, predictive, prescriptive, qualitative)	Profile → clean (with a stated rule) → analyze → validate → report; correlation is never causation, and that rule does not bend under pressure
designing-workflows	You're turning a goal + findings into an ordered, executable plan	Outcome + how it's verified, single-action steps ordered by dependency, a done-criterion per step, parallel markers, checkpoints, and failure paths
orchestrating-tasks	You have a plan and must execute every step and take action	Every step becomes a tracked task; execute → verify the result → then mark done; stop before any destructive / irreversible / outward-facing action

They chain into one pipeline — see docs/research-to-execution-workflow.md:

Research → Analyze → Plan → Execute

Plus a fifth, invisible skill — using-maxforge — the bootstrap that makes the other four reliably auto-trigger. It loads at session start via a hook and tells the agent to invoke the right skill before acting. Without a bootstrap, bare skills fire inconsistently (measured: 2 of 4); with it, all four fire. See the eval report for the numbers.

Quickstart (try it in 30 seconds)

No install needed — load the plugin into a one-off headless session:

claude -p "Analyze this: month,price,units | Jan,10,100 | Feb,12,90 | Mar,15,70. Does raising price hurt sales?" \
  --plugin-dir .

The agent auto-invokes analyzing-data, profiles the data, and refuses to call correlation causation — that's the skill working.

Install (Claude Code)

# Register the marketplace (point at wherever you host this package)
/plugin marketplace add mancaf779-boop/-Max-Geld-Generation

# Install the plugin
/plugin install maxforge-lab

Or, for local development, add the marketplace from your checkout — the plugin lives at the repository root (.claude-plugin/marketplace.json).

The using-maxforge SessionStart hook loads automatically once the plugin is installed, so the skills auto-trigger on matching tasks — you don't invoke them by hand.

How it works

The skills are plain Markdown with YAML frontmatter (name + description). The agent reads the descriptions to decide which skill applies to the task in front of it, then loads the full skill body when it acts. There is no runtime, no dependency, and nothing to configure.

Quality — this was tested, not vibes

These skills were developed and hardened with a RED→GREEN→REFACTOR eval loop (watch an agent fail without the skill, write the skill, close the loopholes), run with real subagents under pressure. Full record in docs/eval-report.md. Highlights:

• ~30 subagent runs across paired baseline/treatment, maximum-pressure, cross-domain, and weaker-model rounds.
• Two skills had real under-pressure failures (giving a false causal "yes"; dropping the confidence/assumption line under "no caveats" pressure). Both were caught, refactored, and re-verified.
• All four held across multiple domains (medical, product, finance, legal, competitive intel, ops, ML, marketing) and on both Opus- and Haiku-class models.

Known limitations (read before shipping a product on these)