AI DIRECTION

Practical AI systems for small products.

Choose narrow, useful automation over spectacle. Define the job, keep the boundaries visible, and make the output reviewable.

Start an AI direction See our work

OUR THESIS

Use the smallest system that can do the job.

Complexity is the enemy of reliability. The useful path has clear limits, visible behavior, and human checkpoints so the decision stays under control.

PROMPT SYSTEMS

Prompt systems that behave like a product spec.

Use this lane when the model itself is not the hard part: the work is defining inputs, allowed sources, voice, output format, examples, review rubric, and version trail for repeated drafts, summaries, classifications, or support output.

PROMPT SYSTEMS Define the contractName the input fields, audience, source boundaries, allowed claims, blocked claims, and exact response shape.Teach with examplesAdd sample inputs, ideal outputs, edge cases, tone examples, and notes for what should be rejected or retried.Evaluate the answerUse missing-field checks, unsupported-claim checks, tone checks, and revision rules so the prompt can be tested.Package the specLeave a copy-ready prompt, usage notes, example set, known limits, changelog, and version trail.

Define the contract: Name the input fields, audience, source boundaries, allowed claims, blocked claims, and exact response shape.
Teach with examples: Add sample inputs, ideal outputs, edge cases, tone examples, and notes for what should be rejected or retried.
Evaluate the answer: Use missing-field checks, unsupported-claim checks, tone checks, and revision rules so the prompt can be tested.
Package the spec: Leave a copy-ready prompt, usage notes, example set, known limits, changelog, and version trail.

Plan Prompting View Work

LOCAL MODELS

Local model workflows for private files and controlled runs.

Use this lane when private source files, hardware limits, model choice, context window, quantization, folder routing, batch volume, or fallback policy decide whether the workflow is viable.

LOCAL MODELS Map the runtimeChoose the local model, model size, hardware target, memory budget, privacy boundary, and hosted fallback rule.Route the filesDefine folder paths, input formats, naming rules, storage limits, backup paths, and what never leaves the machine.Stabilize batchesTrack parameters, context size, seeds where useful, expected output, failure cases, and review queues.Maintain the stackDocument update cadence, model swaps, cleanup steps, fallback models, and the checks needed after version changes.

Map the runtime: Choose the local model, model size, hardware target, memory budget, privacy boundary, and hosted fallback rule.
Route the files: Define folder paths, input formats, naming rules, storage limits, backup paths, and what never leaves the machine.
Stabilize batches: Track parameters, context size, seeds where useful, expected output, failure cases, and review queues.
Maintain the stack: Document update cadence, model swaps, cleanup steps, fallback models, and the checks needed after version changes.

Plan Local Workflow Read Notes

Systems that hold up.

Bounded inputs

Define what goes in, which sources count, and what stays out.

Visible outputs

Make behavior explainable, testable, and easy to review.

Human checkpoints

Keep people in the loop where judgment and responsibility matter.

SYSTEM LEDGER

Direction, runtime, product, and launch in one visible system.

Prompt Systems

Instruction architecture for dependable output: input rules, source boundaries, response shape, examples, and evaluation rubrics.

input contractssource boundariesevaluation rubrics

Local Model Workflows

Private inference pipelines with clear model choice, runtime envelope, file routing, batch handling, and fallback limits.

runtime envelopefile routingprivacy fallback

AI App Prototypes

Turn a rough idea into screens, inputs, outputs, states, and the smallest usable first version.

product flowinteractive UIfirst build scope

Product Direction

Find the useful center of the idea before the build grows extra limbs.

audience fitfeature trimmingroadmap shape

Launch Surfaces

Landing pages, support pages, app store links, privacy notes, and update-list flows that stay current.

web presencesupport pageslaunch copy

Automation And Review

Small pipelines that move files, check outputs, log results, and make human review faster.

batch helpersQA passesstatus checks

METHOD

Ship small. Test hard. Keep control.

Make the useful version judgeable before the idea grows extra limbs.

1Shape
Define the job, audience, input, output, and the decision the tool needs to support.
2Prototype
Build the smallest surface that proves the workflow instead of starting with a giant platform.
3Stress Test
Run edge cases, bad inputs, stale outputs, and review checks before trusting the flow.
4Ship
Package the public page, support notes, tracking, and next-step CTA around the usable version.

CASE NOTES

Useful notes before the build gets bigger.

Prototype BriefWhat goes into a useful AI app prototype brief

Clarify the user moment, inputs, output, review point, and privacy boundary before the first build.

IntakeAn AI workflow intake checklist

Name the user moment, inputs, outputs, review step, privacy boundary, and success signal.

Prompt QAA simple prompt QA checklist

Turn loose prompts into reusable instructions with source rules, output contracts, examples, and rejection checks.

Local ModelsA local model workflow audit

Check model choice, runtime limits, private files, batch settings, review queues, and the final human decision.

FAST ANSWERS

The short version before we build.

01What is the difference between a prompt and a prompt system?

A prompt asks for one result. A prompt system defines the input contract, source boundaries, examples, output format, evaluation rubric, and revision notes so the same job can be checked and reused.

02What makes a local model workflow different?

A local model workflow is about the operating environment around inference: model selection, runtime setup, file routing, hardware limits, repeatable parameters, batch review, privacy boundaries, and fallback rules.

03How small should the first AI app version be?

The first version should prove one user moment: a clear input, useful output, review step, and next action that can be tested before the app grows.

04How do Lowball Lab and Buyer Backup fit?

Lowball Lab is the live secondhand-offer proof app, while Buyer Backup is the proof-packet lane for purchase evidence, claims, warranty tracking, and future Pro exports.

START A BUILD

Need a practical AI path?

Bring the fuzzy version. We will map the smallest system that can do the job and keep the controls visible.

Contact Newman AI Works