BHpresentation · shape of AI
The Shape of the Thing · ten landscape plates · VHX presentation layout
Read essay
BH · Field Note
The shape of AI · presentation
01 Cover

01 Thesis

The shape of the thing.

AI does not arrive as a polished object. It becomes a working system when repeated runs are forced through a visible closed loop.

Format
VHX landscape presentation · 10 plates
Core shape
rectangular loop, not hierarchy
Claim
the artifact is the loop
Heitmann · field note
Possibility needs pressure
01 / 10
BH · Field Note
The common mistake
02 Open loop

02 What fails

Prompt → output is too weak a model for real work.

It can produce language, drafts, and plausible artifacts. It cannot reliably improve without a return path that measures the output and changes the next input.

Promptintention without enough boundary
Model runpossibility, not proof
Outputlanguage exits the system
Open loop · failure mode
No feedback means no accumulated shape
02 / 10
BH · Field Note
Fig. 01 · Rectangular closed-loop AI system
03 Figure

03 Operating shape

The closed system goes around.

This is a loop, not an org chart.

Intention

Goal, constraints, examples, acceptance criteria.

Work

Agent run, tools, files, surfaced artifact.

Memory

Receipts, decisions, context graph, reusable state.

Fig. 01 · closed loop
Loop, not hierarchy
03 / 10
BH · Field Note
Repetition is the material
04 Clay

04 Iterative clay

Every pass leaves a mark; enough marks become structure.

01

Prompt

The initial intention names the desired pressure.

02

Run

The model creates possibility and exposes ambiguity.

03

Inspect

Human, tests, and reviewers find where it deforms.

04

Correct

Criteria, files, prompts, and routes get sharper.

05

Preserve

The next run starts from accumulated shape.

Reps · clay
The tenth pass teaches what the first pass cannot
04 / 10
BH · Field Note
Memory is geometry
05 Graph

05 Memory

A context graph is feedback geometry, not a notes folder.

The durable asset is what survives between runs: facts, decisions, validations, failures, and the edges that explain why they matter.

FactsWhat the system can rely on without rediscovery.
EdgesWhich decisions, constraints, and artifacts caused later work.
ReceiptsValidation evidence, deploy URLs, test output, and review notes.
Context · graph
Memory gives the loop a surface to press against
05 / 10
BH · Field Note
Roles apply pressure
06 Roles

06 Organization

A visible role system makes disagreement cheap.

Context
Graph
PlannerBuilderReviewerWatcher
PlannerTurns intention into a bounded route before tools mutate the artifact.
BuilderMoves the work forward and leaves inspectable changes.
ReviewerChecks the artifact against the contract, not confidence.
WatcherNames drift, stale state, blocked auth, and runtime mismatch.
Roles · loop pressure
Different roles press the same loop from different sides
06 / 10
BH · Field Note
Failure modes
07 Risks

07 Where it breaks

The loop fails when any side becomes invisible.

01No frameThe system works from vibes because intention was never bounded.
02No returnOutput leaves the system and teaches the next pass nothing.
03No memoryEach run rediscovers yesterday's mistakes.
04No witnessThe builder checks itself and calls confidence evidence.
Failure · open edges
Invisible sides become drift
07 / 10
BH · Field Note
Operating contract
08 Build

08 Build principle

Do not buy AI as an object. Build the surface that lets it take shape.

Acceptance before executionDefine the change and the evidence before the run starts.
Tools with receiptsEvery consequential mutation needs a visible test, diff, deploy, or review trace.
Memory after pressurePreserve only the state that actually changed the next run.
Presentation as interfaceDiagrams should teach the operating model, not decorate the page.
Contract · operations
Shape is accumulated constraint
08 / 10
BH · Field Note
Site IA as loop
09 Site
IA · public surface
Route clutter becomes IA
09 / 10
BH · Field Note
Close
10 Ask
The artifact is the loop.

10 Operating ask

Make the system visible before trusting the output.

01Frame intention as a contract, not a wish.
02Force output through feedback before it becomes belief.
03Save only memory that changes the next pass.
04Let the next agent inherit the loop, not just the transcript.
Close · bradheitmann.ai
bradheitmann.substack.com
10 / 10