Lab Intermediate Loop Engineering

The Convergence Room

AI hands you a confident first draft in seconds, so the bottleneck is no longer producing output — it’s converging it to a bar you can defend. This lab is the loop itself: set an explicit standard, then inspect, diagnose, and refine one real draft across rounds, with a live convergence meter that can move both ways. Steer well and it climbs; over-iterate or feed it vague, multi-adjective feedback and you will watch it fall.

6 Steps Run the loop
~30 min Duration
Iterate to a standard Interactive simulation

Step 1 of 6 · Setup

Cold open · A/B

Two ways to run an AI draft

You ask AI to write something. The first draft is not quite right. Here are two operating styles for what happens next. One of them is gambling. Pick the approach you would trust to get you to a defensible result — then check your reasoning.

!

A loop without a standard is gambling. “Regenerate until it feels good” treats AI like a slot machine — pull the lever, stop on a feeling, keep no record of what changed. Loop engineering is the opposite: write the bar first, then steer the draft toward it one measurable change at a time. The rest of this lab is that one move, run for real.

The loop · build & order your standard

Set the standard before you touch the draft

A control engineer never tunes a system on a feeling. They define the target, measure the distance to it, make one adjustment, measure again, and stop when the error is inside tolerance. Loop engineering applies exactly that to an AI draft — six moves, and the first one is the one most people skip.

1

Set Standard

Write the bar before you see a draft.

2

Instruct

Ask for one change at a time.

3

Inspect

Read the output against the standard.

4

Diagnose

Name the specific gap, not a vibe.

5

Refine

Send the single most leveraged fix.

6

Convergence?

Distance closed enough? Stop. If not, loop.

1. Write your standard for the Slack message

Before you touch the draft in the Convergence Room, you set the bar. The artifact there is a blocked-deployment Slack message. Toggle the criteria into your rubric. A standard-strength meter rewards specific, checkable criteria and flags the vague ones as drift-risk — the rubric you build here becomes the scoring frame the simulation uses.

Standard strength 0%

2. Put the loop’s moves in order

The moves are shuffled below. Arrange them into the order a disciplined loop runs. The one thing the lab is most insistent on: the bar precedes the draft — you set the standard before the first instruction, not after you have seen what the model gives you.

    3. Specificity is the steering lever — not round count

    More rounds do not converge a draft; more specific instructions do. Sort each piece of feedback into Steerable (the model can act on it without guessing) or Drift-risk (so vague the model has to invent what you meant, and the draft wanders). Click an item to move it between bins, then check.

    Steerable one unambiguous change
    Drift-risk model has to guess
    Unsorted feedback
    Multi-round simulation · carried state

    The Convergence Room

    Here is the live loop. A 4-line client status email sits at 40% of your standard — it buries the decision and hedges. Each round, pick one instruction. The draft visibly rewrites and the meter moves. It can move down. A targeted, single-change instruction climbs it; a vague multi-adjective instruction reintroduces a hedge and drops it. You are aiming to close the distance to the standard line, then stop.

    Convergence to your standard 40%

    Round 1 of up to 4. Pick the most leveraged instruction.

    Scoring against your rubric:
      Client status email — draft Round 1
      Diagnose & choose the next instruction

      Steer it: diagnose, don’t react

      1. Diagnose this round

      Convergence (carried from the room) 40%

      A “feels off” reaction can’t be sent as an instruction. Locate the exact spans that violate the standard.

      This is the Round-1 draft. Instead of reacting to a global “something’s off,” click the exact spans that violate your standard (buries the decision, hedges, no owner, no date). Naming the span is what turns a vibe into a sendable instruction.

      Click each span you think breaks the standard, then press Check.

      and there are a few things still in motion. , though . once things are clearer.

      2. Choose the next instruction

      Convergence (this branch) 65%

      A different artifact: a job description sitting at 65% of standard. Each instruction has a consequence.

      New artifact, same discipline. A job description is at 65% — the bullets are generic and the title is right. You have one instruction to send. Each branch shows the consequent draft and how the meter moves. Choose, watch the consequence, then try the others.

      Job description — after your instruction

      Call it · pick the signal · spot the failures

      Know when to stop

      1. When has it converged enough to stop?

      Refinement has diminishing returns. The hard skill is calling the stop before you over-iterate a good draft into a worse one. Estimate this: of all the improvement you can realistically get from a refinement loop, what share lands in the first two rounds?

      Drag to your best guess, then reveal the evidence.
      Your guess: 50% of the gains in the first two rounds
      0%50%100%

      2. Match each situation to its stopping rule

      “Stop when it feels done” is not a rule. A deliberate loop names its stop signal in advance. Click a situation on the left, then its correct stopping rule on the right. Link all four, then check.

      3. Spot every anti-pattern in this transcript

      Here is a real-feeling refinement session. It contains several of the classic failure modes at once. Select every anti-pattern you can find, then check — the skill is exhaustiveness.

      You: Write the launch email. (No standard stated.)

      You: Hmm, make it better.

      You: Actually make it punchier and warmer and shorter and more formal and add stats.

      You: Now I want it really casual instead.

      You: (round 7) Still not feeling it… regenerate again.

      You: I’ll just keep going until something clicks.

      Write it · the protocol · reflect

      Takeaway

      1. Write the one instruction you’d actually send

      Back to the Round-1 client email from the Convergence Room. You get one instruction. Write the single most leveraged change — the one that closes the most distance to the standard. Then self-assess against the checklist and compare to a model answer.

      The draft you’re steeringRound 1 · 40%
      Hi team, we’ve been making good progress and there are a few things still in motion. After weighing the options, it may be that the launch could potentially shift, though we’ll know more soon. Someone will follow up once things are clearer. Best, the team

      2. Your loop, on a card

      Capture a reusable protocol you can run on any AI draft tomorrow. This is the discipline you just practised, compressed to six lines.

      My refinement loop

      BeforeI write the standard down — checkable criteria, not “make it good.”
      Each roundI name the single biggest gap and send one specific instruction for it.
      I measuredistance to the standard closing — not whether it “looks different.”
      I neverchange five things at once, or move the bar after seeing the draft.
      I stop whenthe rubric is met, two rounds add nothing, or the gap needs me, not the model.
      I remembermost of the gain is in rounds 1–2; over-iteration can make it worse.

      3. Field guide

      One page to keep. Everything the Convergence Room taught, at a glance.

      1The loop

      • Set Standard → Instruct → Inspect
      • Diagnose → Refine → check Convergence
      • The bar precedes the draft

      2Set the standard first

      • Checkable criteria only (“states the blocker in one line”)
      • “Sounds urgent” / “is professional” are drift-risk
      • Never move the bar after seeing output

      3Feedback template

      • One change per round
      • Name the span / fact / structure
      • Specificity, not round count, is the lever

      4Diminishing returns

      • Most achievable gain lands in rounds 1–2
      • Near-plateau by round 3
      • Over-iterating can lower the result (drift)

      5Three stop signals

      • Meets-rubric — bar cleared, ship
      • No-improvement — two flat rounds, plateau
      • Escalate — judgment call; you take it over

      6The danger

      • A loop without a standard is a slot machine
      • “Looks different” ≠ “closer”
      • Vague, multi-adjective feedback reintroduces hedges

      You’ve run the loop.

      You set a standard, steered a draft to it one change at a time, watched the meter fall when the feedback went vague, and called the stop before over-iteration cost you. That is the difference between using AI as a slot machine and using it as a controllable instrument.

      Back to the Lab overview
      Step 1 of 6