In developmentfromScratch
Project management where AI plans and you decide.
fromScratch gives a team a network of AI experts that decompose plans, research and coordinate, while every strategic call and approval stays with a person.
No mail app? Write to contact@nsi4.com
- Project. The work the canvas is for, as one node.
- Decomposing. Items the AI is still breaking down, and how many are left.
- Plan. Each inferred item, with the share of it that is done.
01 Problem
One prompt at a time does not run a project.
Management that can't keep up
Status chasing and admin crowd out the work. Teams sit in silos, plans come out too detailed or too thin, and the tools buckle as the project grows.
The single-prompt trap
One question, one answer, nothing carried forward. Results vary from run to run, and everyone starts from scratch every time.
Tools that don't talk
GitHub, Jira, HubSpot and Slack each hold part of the project. Information is duplicated, nothing syncs, and knowledge falls into the gaps.
Each problem makes the other two worse.
02 Constraints
People keep the final say.
- Authority
- Strategy, creative decisions, approvals and stakeholder relationships stay with people
- AI's role
- Planning, decomposition, research and coordination, under a person's direction
- Providers
- Works with more than one AI provider
03 What We Built
A network of AI experts, not a single chatbot.
Four experts work every plan through one prompt framework, and nothing applies until a person approves it.
How a plan moves
- DirectArchon, the orchestrator, directs the other experts on the plan
- CurateNexus gives each expert the context it needs and filters out the rest
- WorkSpecialists work one item at a time, under a person's supervision
- FlagOracle flags risks before they land
- DecideNothing applies until a person approves it
The experts
Archon
The orchestrator. Directs the other experts on a plan.
Nexus
The context curator. Gives each expert the context it needs and filters out the rest.
Oracle
The predictive guide. Flags risks before they land.
Specialists
Task-focused experts that work one item at a time, under a person's supervision.
The Codex
Sixteen prompt categories in six groups, so every expert's instructions are built the same way.
Identity and direction
- System persona and role
- Overall goal
- Primary task
Context
- Domain background
- Input to process
- Target audience
Guidance
- Worked examples
- Reasoning method
Output
- Format and structure
- Tone and style
- Quality criteria
Boundaries and safety
- Constraints
- Ethical guardrails
- Fallback and clarification
Technical control
- Delimiters and section markers
- Temperature
A professional library of system personas, output formats, task templates and quality standards comes with it, and a team can add its own patterns and knowledge.
The platform
One project record
Sync with GitHub, Jira, HubSpot and Salesforce, so the project lives in one place. Planned, not yet built.
Collaborative planning
The experts draft the plan, break it into tasks and track dependencies. A person approves.
Any AI provider
Each task goes to the provider that fits its quality bar and its cost.
Help in context
Suggestions and risk flags across every feature, applied only when a person accepts them.
How it thinks
Context like memory
The HCLC context system is modelled on how human memory activates: it brings forward what matters and filters out the rest.
AI that reviews AI
Experts check and refine each other's output in rounds, instead of stopping at the first answer.
Experts that collaborate
Agents hand work to each other under set rules for collaboration and disagreement, and report to the person leading.
04 Evidence
Every AI change waits for a person.
The approach comes from NSi4's own research on AI multiplication and prompt complexity.
- Items. Each line of the quote the AI turned into work.
- Progress. The share of each item it judges done.
- Review. The ✓ and ✗ beside an item accept or reject it.
Oracle · proposed change
Illustration
Bakery website › Online ordering
Why. Card payments need a payment provider account, and no item in the plan opens one.
05 Stack
What it runs on.
- Status
- In development
- AI
- Multiple providers, matched to the task
- Research
- AI Multiplication · Prompt Complexity
- Built with
- .NET · SQL Server · multi-model AI
Early access
Plan your next project with fromScratch.
fromScratch is in development, and early access opens to a small number of teams. Tell us about your team and the projects you run.