Most AI workflow advice starts with tools.
LADDERGATE starts with the work.
LADDERGATE is a folder-based AI researcher for builders designing Claude project folders, skill stacks, Graphify graphs, SQLite state layers, slash-command workflows, and agentic production systems.
Every AI system you could build sits somewhere on this ladder:
PROMPT → FOLDER → SKILL → COMMAND → AGENT → STATEFUL AGENT → GRAPH-ASSISTED SYSTEM
Each rung is justified by real requirements — not by hype, tool availability, or engineering preference.
LADDERGATE's job is to stop you climbing this ladder before the workflow earns it.
Give LADDERGATE a workflow idea. It investigates before it prescribes:
- Should this be a prompt, folder, skill, slash command, agent, database, or graph?
- What state actually needs to persist between runs?
- Where will context contamination occur?
- Is Graphify useful here, or will it make the system worse?
- When is SQLite justified?
- When is a "multi-agent system" just overbuilt prompt chaining?
LADDERGATE refuses to recommend tools until it has classified the workflow, the state requirement, the knowledge boundary, and the failure cost.
Before prescribing architecture, LADDERGATE runs every workflow through this gate:
- Task type — one-off, repeated, deterministic, long-running, or multi-agent?
- State requirement — no state, temporary context, folder memory, structured database, or external source of truth?
- Knowledge boundary — where does the truth come from?
- Failure cost — harmless / annoying / expensive / reputational / legal-medical-financial?
- Architecture class — which rung on the ladder is the minimum justified answer?
Every LADDERGATE recommendation is challenged by the Drift Auditor — a hostile internal reviewer that attacks the proposed architecture before the builder commits to it.
The Drift Auditor asks:
- Does this actually need to persist, or are you assuming it does?
- What happens when the agent loses context mid-run?
- Could a well-written prompt replace this entire system?
- What is the failure mode if the graph is stale?
- Are you building an agent because the workflow needs one, or because it sounds powerful?
- Drop this folder into your Claude project
- Describe your workflow idea
- LADDERGATE classifies it using the five-stage gate (
rules.md) - LADDERGATE prescribes the minimum justified architecture
- The Drift Auditor reviews the prescription
- You receive a build path — not a hype stack
| File | Purpose |
|---|---|
identity.md |
Who LADDERGATE is, what it will and will not do |
rules.md |
Five-stage classification gate + Drift Auditor protocol |
examples.md |
Four annotated decision scenarios |
reference/assessor-guide.md |
Judge-friendly 5-minute test path |
reference/architecture-decision-ladder.md |
Full ladder with justification criteria per rung |
reference/failure-patterns.md |
How workflows fail when architecture is skipped |
reference/source-hierarchy.md |
Evidence tier list and claim tagging system |
reference/folder-vs-agent-vs-database.md |
State taxonomy for decision-making |
tests/ |
Four scenario test files matching the demo cases |
| Factor | LADDERGATE's position |
|---|---|
| Specificity | Not "AI researcher" — "researcher for agentic workflow architecture decisions" |
| Usefulness | Solves real pain: people overbuild, contaminate context, create fragile systems |
| Folder-native design | The system's entire operating model matches the challenge format |
| Judging clarity | Assessor guide lets a judge test it in 5 minutes |
| Unfair advantage | Can help people build better researchers — including this one |