Skip to main content
Approach

How the work moves

This is the loop I run on code changes, from the first idea to the deploy. Each step belongs to one of the six seats. Three of them are gates that a person has to clear.

  1. Step 1

    Write down the intent

    Business Analyst

    Before any work starts, the idea goes on a ticket: what is being built and why. Nothing gets built without one.

  2. Step 2

    Challenge the idea

    Business Analyst

    The idea is questioned on its value, its return and the five W's. It gets reshaped or killed here, and doing nothing is always one of the options.

  3. Step 3

    Plan it and set the checks

    PMO

    The plan names the files, the approach and what done means. Each acceptance criterion becomes a row in a checks table that starts unresolved, and each step of the plan gets an autonomy level.

  4. Step 4

    Own the plan

    Principal ConsultantHI gate

    I explain the plan cold, in my own words, before it is approved. If I can't, it isn't approved.

  5. Step 5

    Build to the plan

    Developer

    The code is written to the agreed plan in house style. Anything outside the plan waits for its own ticket.

  6. Step 6

    Challenge and simplify the code

    Developer

    The diff is read the way a hostile reviewer would read it, then cleaned up: reuse what already exists, cut dead code, pick the simpler form.

  7. Step 7

    Own the code

    Principal ConsultantHI gate

    I explain how the change works and why it is written this way before it leaves the branch.

  8. Step 8

    Open the pull request

    PMO

    The pull request carries the checks table, refreshed on every push and naming the commit it was checked on.

  9. Step 9

    Review in a separate session

    Code Reviewer

    A session that didn't write the code reads it with none of the builder's context, runs the secret and security scans, and grades every finding from Critical to Low.

  10. Step 10

    Test end to end

    QA

    The end-to-end suite runs against a preview of the exact commit that will merge, never against production. TestProof turns the run into a report someone who doesn't write code can follow.

  11. Step 11

    Go round again if it isn't clean

    PMO

    A Critical or a major finding sends the change back to the code challenge on the new head. A failed check gets three tries, then it comes to me as a blocker.

  12. Step 12

    Merge on green, then ship

    Principal ConsultantHI gate

    Green means an approved review, passing tests and a full checks table, all on the same commit. In Dispatch, every merged pull request adds an entry to the app's What's new page, and a test fails if one is missing. A merge that deploys waits for my word.

Under every step

  • House rules and specialist agents

    Written house rules load into every session, and specialist agents take the narrow jobs.

  • Sensitive-data scan

    Every tool call is scanned for sensitive data. The scan logs what kind it saw and how much, never the data itself.

  • Injection scan and quote check

    Text read from outside is scanned for prompt injection, and quotes are checked against what was actually read. CopernicIQ, the decision agent on the Spark, judges both.

The systems it runs on

  • GitHub: code and pull requests
  • Vercel: previews and deploys
  • ClickUp: the intent ticket
  • Notion: the planning record
  • Dispatch: our own project tool
  • The Spark: local models and CopernicIQ
  • E2E hub: the end-to-end suites
  • TestProof: readable test reports

The HI Script

HI means human intelligence. An HI Seat is filled by a person and an AI Seat by an agent, and the charter is the same either way. The HI Script is the rule that decides which, seat by seat, on each engagement. It is four questions, and each one is a rule I already work by.

  1. Test 1

    Who signs?

    The person who signs for a change has to be able to explain it, and releases and deploys wait for a person's word. Signing is always HI.

  2. Test 2

    What breaks if it's wrong?

    If a step touches money, security, production data or anything hard to undo, a human gate stays on it. That gate can't be put off or skipped.

  3. Test 3

    What proves it?

    An AI Seat runs a step on its own only when tests fully cover that step. If nothing would catch the mistake, a person checks it.

  4. Test 4

    Has it earned it?

    Autonomy goes up one level after three clean runs in a row, and a person decides. One failed run resets the count.

Today the Principal Consultant seat is HI. The other five run as AI Seats behind HI gates.

Built Something with AI?

I'll spend 30 minutes telling you what my production-readiness review would flag.