FAQ¶
What is AgentCraft?¶
A curated, open-source, MIT-licensed library of three artifact types — scripts (deterministic, no LLM), skills (agent-loaded procedures), and flows (multi-tool compositions) — built from 9 years of O&G and API-auditing work. Each artifact is published with a trigger prompt, a one-line install, a sample output, and a last_tested date.
Is it free?¶
Yes. MIT-licensed, no paid tier. Some flows use paid third-party APIs (Apify, DeepSeek, Gemini, etc.), but each run is <$1, and the offline scripts run with zero API cost. The site is static HTML, hosted on a free tier of Cloudflare Pages or Vercel.
How is it different from the 240+ Hermes skills I already have?¶
Hermes is the engine. AgentCraft is the curated, public, O&G-specific catalog built on top of it. We promote 20-40 high-quality, O&G-relevant artifacts from the underlying skill set, wrap them with source attribution, trigger prompts, and a public site. The other 200+ stay in your private Hermes.
How is it different from the Claude Code plugin ecosystem?¶
Claude Code plugins are installable modules with slash commands, agents, and hooks. AgentCraft is a library of recipes — the trigger prompt is the unit of value, not the install. A visitor to AgentCraft copies a sentence, pastes it into their agent, and gets a real artifact. The install is implicit; the value is the recipe.
What does "domain-first" mean in the catalog?¶
Drilling / Quality / HSE / Operations is the primary navigation. You browse by what you do, not by what tech the artifact uses. Type (Script / Skill / Flow) and vertical (Upstream / Midstream / etc.) are secondary filters. A fallback "use-case-first" navigation exists for cross-cutting workflows (audit, training, tendering).
Can I submit my own artifact?¶
Yes, PRs welcome. Every artifact must:
- Have a source-of-truth document in the Drive
00-canon/folder - Include a trigger prompt that loads the right skill when pasted
- Be reproducible (same input → same output, or variance documented)
- Cite the standards it uses correctly (API Q1 11th ed., not "Q1"; ISO 9001:2015, not "ISO 9001")
- Pass
CONTRIBUTING.md(filename convention check)
How do I know an artifact is still good?¶
Every page shows a last_tested date in the footer. CI runs every flow weekly (Mondays at 06:00 UTC) with a 3-5 minute smoke test. A failure sends a Telegram alert to the maintainer. Stale artifacts (no test in 90+ days) are flagged with a yellow banner; broken artifacts (test failing) are flagged red.
What if a flow breaks because an API changed?¶
- The flow's
CHANGELOG.mdis updated within 24h of the fix - The public page shows a "broken" banner until the fix lands
- Drive's version history lets you roll back to the last working version
- The maintainer publishes a
vN.N.(N+1)patch in the same week
Why is the catalog only 3 artifacts at launch?¶
Because curation is the value. A library of 240 artifacts with no curation is a search engine, not a library. The 20-40 artifacts that earn a place in AgentCraft will be the ones that:
- Solved a real problem for a real client
- Have a source-of-truth document in Drive
- Have a trigger prompt that consistently loads the right skill
- Are tested weekly without flakes
Day-7 target: 7 artifacts. Day-30 target: 20. Day-90 target: 40.
What's a "trigger prompt" and why does it matter?¶
A trigger prompt is the literal sentence a visitor pastes into their agent to invoke an artifact. Example for the DEPRO calc:
Run depro_calc.py on inputs.yaml for Well A-12.
The trigger prompt is the secret weapon of AgentCraft. It turns a static library into an immediately usable one. A visitor doesn't have to guess how to invoke a skill; they copy a sentence, paste it, and the agent does the right thing.
Every artifact page has a "Trigger prompts" section with 3 variations. Copy any.
Does this work on Windows / Linux / macOS?¶
Yes for scripts (pure Python, stdlib). Skills and flows need an agent runtime — Hermes works on all three, Claude Code works on all three. The site is pure static HTML, deploys anywhere.
How do I deploy the site myself?¶
To Cloudflare Pages:
To Vercel:
Both free tiers work.
Where's the source brief?¶
agentcraft-strategy-brief.md in the repo root. 763 lines, 16 sections, the full strategy this library was built from.
I have feedback or a request.¶
Open an issue on GitHub. Use the enhancement or request template. Be specific about your O&G / QMS / API-auditing use case — the maintainer prioritizes by frequency × impact.