How it works¶
AgentCraft has three artifact types. The decision of which to use is the first thing to learn.
The decision tree¶
Does the task need an LLM to interpret, decide, or write?
├── NO → Script. A single file. Deterministic. Zero API cost.
└── YES
├── Does it need only ONE tool/API and <5 steps? → Skill.
└── Does it need 3+ tools, 5+ stages, multi-output? → Flow.
The three types¶
Script¶
A single, deterministic, copy-pasteable file (Python, Bash, SQL, JS) that does one job. No LLM in the loop. Same input → same output, always. No API cost. Usually stdlib only.
| Example | Domain | What it does |
|---|---|---|
| DEPRO Calculator | Drilling | Calculate Daily Equivalent Production Offline + Monte Carlo uncertainty |
| API Q1 Audit Checklist | Quality | Generate 35-clause audit checklist with severity + gap detection |
| Pareto chart | Operations | CSV → ranked bar chart with cumulative % line |
| Fishbone diagram | Operations | Free-text causes → structured Ishikawa SVG |
Skill¶
A SKILL.md with YAML frontmatter that an agent loads to gain a repeatable procedure + the right tools/env to execute it. The skill teaches the agent how to do a specific thing, in your voice, with the right tool calls.
| Example | Domain | What it does |
|---|---|---|
skill-iwcf-l4-drill |
HSE | Run a 30-min IWCF L4 well control drill — agent plays examiner, you play candidate |
skill-cement-job-design |
Drilling | Produce a cementing job design from a well spec |
skill-audit-non-conformance |
Quality | Review an NCR against API Q1 11th ed., output a gap list with severity |
skill-depro-calc (wrapper) |
Drilling | Wraps the DEPRO script with the right voice + checklist + chat output |
Flow¶
A multi-stage, multi-tool composition (an agent + a sequence of skills + scripts + APIs) that produces a complete deliverable. Flows are the closest thing AgentCraft ships to "an entire product in one run."
| Example | Domain | Stages | Cost / run |
|---|---|---|---|
| SME Audit Deliverable | Operations | 8 stages: intake → BMC → PESTEL → gap → roadmap → vendors → package → delivery | ~$0.54 |
| Lead Research | Sales | ICP → Apollo/Apify discover → DeepSeek enrich → Airtable | ~$0.55 |
| Content Publish | Marketing | Topic → Exa research → DeepSeek draft → Gemini visuals → Beehiiv newsletter | ~$0.29 |
| Build Website | Web | Brief → IA → copy → design system → visuals → Next.js code → Vercel deploy | ~$0.17 |
The three filters visitors use¶
When you browse the library, you can filter by:
- Domain — Drilling / Quality / HSE / Operations
- Type — Script / Skill / Flow
- Vertical — Upstream / Midstream / Downstream (drilling); API Q1 / API Q2 / ISO 9001 (quality); IWCF / IADC (HSE); LSS / SME Audit (operations)
Plus a tag overlay (#offline, #needs-api, #abu-dhabi, etc.) for cross-cutting filters.
What's NOT here¶
- Generic productivity skills (todos, calendars, email) — those belong in Hermes, not here
- Skills with no source-of-truth document in Drive
- Skills that only work on one specific client's data
- Anything where the LLM does the heavy lifting and the input/output isn't a real artifact
Why domain-first?¶
A petroleum engineer asks "do you have anything for cementing?" not "do you have any skills that take a YAML well spec and produce a 12-section markdown report?" Domain-first matches how visitors think. The type filter is secondary because visitors care that the answer exists, not what tech it uses.
Full taxonomy: agentcraft-strategy-brief.md §3.