Workspaces you can download and use today.

A small business process often works because the owner is holding it together. The rules live in their head, the exceptions are handled from experience and everybody knows who to ask when something does not fit. The problem appears when the work is handed to an assistant, a new employee or an AI. What felt obvious was never written down.

Each workspace takes one of those jobs and makes the working method visible. It defines what information is needed, what happens next, what should stop the process and where a person needs to make the final decision.

These are working demonstration systems, not screenshots or speculative concepts. You can download them, run the tests and follow a piece of work from beginning to end. Adapting one to another business means replacing the example rules and business details, while keeping the tested process underneath.

The work stays in ordinary files that you can open, read and change. There is no dashboard concealing what happened or why. Open the workspace and you can see what has been completed, what is waiting and which decision moved it forward.

That is the evidence here. You can inspect the thinking, test the boundaries and see whether the system holds without taking my word for it.

The method is Interpretable Context Methodology, published byJake Van Clief and McDermott, MIT licensed. These are my builds on it.

Built for: the consultant, coach or small business owner who uses AI to help draft content but whose own voice is part of why clients trust them.

The problem: AI makes the first draft faster, then quietly replaces your language with its own. You either publish something that sounds like everybody else or spend the saved time rewriting it yourself.

Author Voice Checker reads the draft line by line and identifies where the writer’s voice has drifted into machine cadence. Each finding names the rule that was broken and explains the problem. It deliberately does not rewrite the copy, so the writer keeps control of the voice.

What it proves: a working editorial system built in two forms, one using deterministic code and one using an AI model, with evidence showing where each approach succeeds and fails. It turns subjective editorial judgement into specific, testable feedback without pretending the final decision can be automated.

Built for Clief Notes competition 9. Honourable Mention.

View the public repository.

Built for: the owner of a small business hiring their first copywriter, marketing assistant or VA.

The problem: your customers recognise your voice because, until now, you have written everything yourself. The first person you hire sends back polished copy that could have come from any business. You rewrite it, explain the same preferences again and eventually decide delegation is more trouble than it is worth.

If somebody else is about to write under your name, this gives them a voice guide they can use before drafting and a mechanical check that catches named rule breaks before the work reaches you. It does not pretend a score out of ten can tell you whether writing is good. It shows the writer which rule failed, where it failed and leaves the repair with them.

What it proves: a three-stage system for briefing, checking and handing copy back to a writer. The voice rules exist in both readable guidance and working code, with checks that stop the two drifting apart. The live self-test passes 28 checks, including known good and deliberately bad drafts. The owner's final read remains part of the process.

Built for Clief Notes competition catch-up. Not submitted. All examples use the fictional Ruff Cuts mobile dog-grooming business.

Built for: the solo financial planner, accountant, broker or adviser who spends the opening of every first meeting establishing the same basic position.

The problem: the first twenty minutes disappear into questions and arithmetic before the useful conversation can begin. Some prospective clients never book because they are frightened of what the numbers might say. A calculator could save the time, but a bare shortfall only tells somebody they have a problem and leaves them alone with it.

Where You Stand asks for five figures and returns one position: behind, on track or ahead. When it finds a gap, it must also show quantified ways to close it. That requirement is enforced in the code. The result appears before any request for an email address, and every assumption used in the calculation is shown.

What it proves: I can take an expert's repeated first-meeting reasoning and turn it into a standalone browser tool without flattening it into a number. The calculation model, browser version and command-line version are checked against one another. The live self-test passes 51 checks, including the rule that a shortfall can never appear without a path attached.

Built for Clief Notes competition catch-up. Not submitted. This is a demonstration using a simplified retirement model, not financial advice.

Built for: the business owner bringing in their first assistant, VA or part-time admin help.

The problem: everything lives in your head and three different documents. Your new assistant cannot tell a real buyer from a friendly enquiry, so either they ask you about everything, which is slower than doing it yourself, or they guess, and you find out a week later.

If you are about to hand work to someone for the first time, this system defines what they need before they can start, not what you meant when you asked. Each handover is written from the receiving person's side: what must be present, what may be accepted and what stops the work immediately. A missing qualification decision cannot quietly become a drafted reply, and the assistant can never move a draft into a sent state.

What it proves: a working five-stage pipeline moving an enquiry through intake, qualification, context research, owner-voice reply drafting and commitment capture. The handover documents and the code are checked against the same model. The live self-test passes 73 checks, including safe packets, missing information and an assistant attempting to send without the owner's approval.

Built for Clief Notes competition catch-up. Not submitted.

Built for: the business owner or subject expert running a webinar, workshop, training session, demonstration or community call.

The problem: knowing your subject does not mean you know how to hold a room. You prepare more slides because that feels controllable, under-prepare the interaction, ask a question that gets silence and decide the audience is not interested. Generic advice about icebreakers will not tell you what to do with twelve willing people on camera or eighty compulsory attendees after a long meeting.

The Room Coach refuses to give you a technique until you have described the room: how many people, whether they chose to attend, what they can do, what happened in the hour before yours and what the person paying needs them to leave able to do. It then gives you one risk, one interaction loop, one rehearsal and one push for that room.

What it proves: I can turn twenty-seven years of live training judgement into a system that asks for the right context before it responds. Eight room facts drive a bounded set of coaching positions, with tests showing that each fact can change the result for a reason. The live self-test passes 212 checks and blocks vague rooms, generic technique lists and claims that a session went well without evidence.

Built for Clief Notes competition catch-up. Not submitted.

Built for: the business owner, bookkeeper or small-practice accountant reconstructing a period from bank lines, partial paperwork and memory.

The problem: incomplete records can still produce a tidy-looking total. That is what makes them dangerous. A recollection, an inferred fee and a bank entry can end up beside one another in the same spreadsheet, with nothing showing which figure can carry weight and which one still needs evidence.

Paper Trail gives every figure a source class and makes the missing-record list the main output. It ranks what needs chasing by the effect on the total, the risk of getting it wrong and whether the missing source can be found. Recollection and inference can explain a gap, but neither can support a stated total. Conflicts remain visible and the reconstruction cannot call itself complete while unresolved items remain.

What it proves: a four-stage research pipeline that inventories sources, classifies evidence, ranks gaps and produces a reconstruction report without quietly filling holes. The live self-test passes 99 checks, including unsupported totals, hidden discrepancies, false completion and changes that would make the written source rules drift away from the code.

Built for Clief Notes competition catch-up. Not submitted. Every worked example is synthetic. No real business, counterparty or figure appears in the build.

The repos above are the ones I can hand over.

This is the one I can't, because it runs my three businesses and it holds everything in them.

It is a personal AI operations layer, running daily. It schedules jobs, sends real notifications, and blocks its own mistakes with rules written into code rather than left as good intentions.

Thirty-one of those rules are enforced by hooks that stop the work rather than warn about it. Sixty-four test files cover the system. Twenty-one scheduled jobs run unattended every day, including a health check at six every morning that looks at eighteen separate points of failure across email, calendar, backups and infrastructure, fixes what it can safely fix, and names the rest.

That last distinction is the whole job. Unattended automation has to know the difference between "fix it" and "tell someone". Get it wrong and the automation becomes the thing that quietly breaks everything.

I can't give you this one, so I'll show you it instead. Ask and I'll walk you through it running.