01Skilld Design · Supporting capability

Repeated work, made useful.

This is not prompt familiarity. It is defining the real task, its inputs and rules, deciding where a model, deterministic code or a human decision belongs, testing the result honestly, and carrying the system into something a person can actually use.

Brand identity, packaging, websites and ecommerce remain the lead of this practice. AI is a serious third capability with its own evidence, not the first promise.

“I turn repeated, rule-heavy work into useful tools and workflows with clear human review.”

The method

The same six decisions, every time.

  1. 01

    Task

    Define the real repeated task, not a generic category of work.

  2. 02

    Inputs and rules

    Set the exact inputs, rules, exceptions and failure states before any model runs.

  3. 03

    Model / code choice

    Decide where a model, deterministic code or a human decision belongs.

  4. 04

    Quality checks

    Build checks that protect accuracy, cost and output quality before release.

  5. 05

    Human decision

    Keep one person responsible for the judgement a model cannot own.

  6. 06

    Usable result

    Carry the system into a workflow, interface or handover that actually gets used.

Anchor 01 · Visual centre

CueWright

Working internal buildSelf-directed, local-first, used internally.

A native macOS build that uses a chosen translation model to turn a plain-language creative brief into a structured, target-ready cue across ten workflows.

10Workflows
50Dated target-capability records
1,222Documented tests
125Test suites
NoneOwned backend or telemetry

Anchor 02

Anonymised service-business content-production system

Live operating workflowOperating workflow for an established service business.

An approved content request is routed through strategy, copy, creative and artwork. Actual-pixel review, deterministic checks and an independent blind critic gate every release before it reaches a fail-closed state or an exact-hash release record.

4Routed production stages
2Independent review passes per release
HashExact release record, not a filename
CappedHuman-approved correction memory

Anchor 03

Private reconciliation workflow

Private recurring workflowConfidential operating context.

Two drifting spreadsheet inputs are translated into deterministic matching, exception and bank-tie-out rules, producing a review workbook. AI assisted the discovery and design of the rules; the recurring run is deterministic code, not a model call.

2Drifting spreadsheet inputs
3Rule types: match, exception, tie-out
1Human final check before close

How Neil tests AI claims

Built, compared and measured.

Neil evaluates his own systems with the same discipline he brings to client work: a defined method, a controlled sample and a result that informs the next build decision.

Knowledge-Backed Agents · first evaluation 48 sessions · scored out of 24
No policy context
16.83
Raw policy
22.08
Structured knowledge
23.17
Knowledge + workflow
23.25
Swiss-poster test · second evaluation 15 sessions · scored out of 24
Role prompt
23
Curated design knowledge
23.2
Knowledge + workflow
23.2

The curated-knowledge conditions held the same high score; adding workflow instructions did not improve the result.

Internal operating method

The discipline he runs his own practice on.

Internal tool in active use

Skilld Design Hermes

An internal operating system with a closed capability allowlist, a router, seven specialist modes, and defined consultation/write boundaries, running against a verified live profile.

Internal tool in active use

AppleDevKit

An internal Claude plugin of 41 skills with state and checkpoint gates and real-pixel Simulator critique, held to a verified test suite.

Internal tool in active use

MyNoteOrganiser

A local-first pipeline that turns selected Apple Notes into a reviewed knowledge vault, with immutable capture and explicit human promotion before anything is treated as fact.

Selected working-build ledger

Selected self-directed builds.

These working builds show the range of systems Neil has designed and carried into software: native apps, deterministic engines, cost-aware AI services and reviewable workflows.

Start with one process

Describe one repeated, rule-heavy process worth examining.

A focused first conversation to map the task, inputs, exceptions and person responsible for the final decision.