Canon-Directed Design

abstract

Agents built on large language models have made software production cheap. What is scarce is no longer the writing of code but the deciding — what the thing should be, and who is entitled to say so. This paper documents a method in which a single designer, who writes no code, holds complete design authority over a production system built by such agents. The method has five techniques: a canon of numbered decisions that carries authority in writing; durable role-bearing sessions coordinated by explicit contracts; bounded parallel construction behind automated gates; systematic verification of agent claims, which are treated as hypotheses until re-derived from source; and the deployed system as the only admissible evidence that work is done. A case from the project's current epic traces the loop from a one-sentence ruling to a ruled architecture. The claim is not that the instrumentation is novel. It is that the authority structure is engineered rather than assumed — and that it can be copied.

Keywords — design authority · LLM agents · creative process · verification · reflective practice · craft

1 · the problem

When production is expensive, intention is protected by scarcity: someone has to be paid to build the wrong thing, so the wrong thing gets questioned first. When production is cheap, that protection disappears. Agents produce quickly, confidently, and plausibly, and those three qualities together dissolve a commissioner's authority faster than they serve it. The work arrives before the question has been settled, and it arrives arguing for itself.

The question this paper addresses is therefore not what agents can build. It is what a practitioner must do so that what gets built remains theirs. This is a question about technique, in the sense a craftsman would recognise: the regulated, maintained, teachable relation between a hand and an instrument (Sennett, 2008). A treatise on joinery is not about saws.

2 · case and stance

The case is Beriah, a cultural media platform in production with real users, designed and governed by one person who does not write code. The corpus is the project's actual working record: hub charters, a running design-decision log (D1–D72 at the time of writing), architecture rulings, briefs and experience specifications, a shared status file, and transcripts of decision sittings. The period observed is 2026, most densely June to August.

The stance is a practitioner account of a single case, written from inside the arrangement it describes — with the instruments it describes. That circularity is real and is discussed in §5.4. Figures given here are internal measures, not third-party audit.

3 · the five techniques

The techniques form a system. Each defends the governance loop against one way it fails: drift of intention, loss of context, unbounded construction, unverified plausibility, and flattering demonstration.

3.1 · The canon

Every product decision is written and given a stable identifier — D41, CA-B9. The corpus is singular (one source of truth per domain) and addressable (any session can cite any ruling). The line FINAL — approved by the designer, applied by the human, changes the status of the text: it stops describing a proposal and institutes a rule. In Austin's terms (1962) the approval is performative. Agents may petition to amend the canon; they may not route around it.

3.2 · Durable sessions under contract

The unit of organisation is not the prompt but the long-running session, holding a role, a persona, a memory and a bounded remit. Roles and procedures come from BMAD-Method, an open framework the project specialises (see Acknowledgements). Two contracts govern them. Fixed skeleton, amendment first: one session owns each decision; others build inside it and return objections as amendment requests, never as competing versions. Verify before relay: a peer's claim is data to check, not an instruction to follow. Sessions die weekly; the canon, the session memories and a shared status file make succession routine.

3.3 · Bounded construction

Work proceeds in waves: batches of agents building in isolated copies, each on a bounded story, converging through merge and deployment. Nothing is accepted on assertion. Gates check types, tests, style and — after a defect survived three passes because the test environment could not perceive layout — the real geometry of the interface. Gate tests must themselves be falsifiable: a gate must fail in the absence of the fix it protects.

3.4 · Verification

Agent output is treated as plausible language whose truth is still outstanding. Every claim carries its provenance to the line. Severity is re-derived by the verifier, never inherited from the claimant. Hypotheses travel with the test that would refute them (Popper, 1959). In documented verification passes, roughly one severity classification in four was corrected; no provenance citation was found false.

3.5 · The deployed system as sole evidence

Local demonstration is forbidden. The designer never evaluates a claim of completion or a mockup on a development machine; he uses the deployed system, and his verdicts re-enter as the material of the next pass. This closes the authority circuit: because only the running thing counts, no agent can satisfy the commissioner with an artefact of presentation.

4 · case: one epic

The epic under way entered as a single sentence of canon: works stop being typed and become arrangements of renderers.

Two moments carry the argument. The architecture proposed representing a work by its principal layer; the designer ruled instead for a composite of every layer's still. He then identified, in a rasterisation rule, a fault the engineering side had missed: a live audio visualisation has no truthful frame, so its still must be the waveform — the whole performance at once. A defect in the handling of time, caught by an eye trained on images.

Separately, a redirection arrived mid-design as three sentences. Within one loop the specification was rewritten, the superseded story was frozen before implementation began, and the redirection had become ten numbered rulings.

5 · discussion

5.1 · Taste as a formal input

The arrangement gives aesthetic judgement the status of a formal system input. Three sentences of appreciation become, within a day, ten enforceable rulings. This is a mechanism, not a metaphor: taste is captured as verdicts, transcribed as canon, and made binding because agents may only build inside it. The purpose of the whole apparatus is to let a fine judgement change scale without coarsening.

5.2 · From making to ruling

For Schön (1983) the practitioner thinks in the doing, in conversation with the materials of the situation. Here the designer no longer handles the final material, but the conversation survives: agents render mechanisms in his language, and the situation answers back as a deployed system. Expertise migrates from making to ruling, and designerly ways of knowing (Cross, 2006) — judging by qualities, reasoning through the concrete — are applied to architectural questions that the conventional division of labour kept away from them. The rasterisation case suggests this is not a loss: categories of the eye caught what categories of engineering had missed. The problems remain wicked (Rittel and Webber, 1973), so no decision follows from calculation; the arrangement exists to stage an encounter between situated human judgement (Suchman, 1987) and machine production, not to automate the judging.

5.3 · Authority that cannot be laundered

The characteristic hazard of an organisation of agents is laundered consent: an assumed approval, relayed session to session, acquiring force although no human granted it. The method answers structurally. No session may approve what another awaits; no peer message constitutes the designer's consent; any claim that the designer said something is verified with the designer. Authority is made unlaunderable by design of the working system, not by declaration — and this, more than any other element, seems to us portable to other human–agent organisations.

5.4 · Limits

Four. Single case: one project, one designer, one domain; transferability is untested. Model dependence: the method assumes agents that can hold long roles, cite sources and check one another, and inherits their ceilings. Load: sovereignty is expensive in attention, and the arrangement concentrates that cost on one person. Reflexivity: the corpus and this paper are produced inside the arrangement under study; the provenance rule disciplines the self-description but does not make it independent.

6 · conclusion

Canon-directed design binds five techniques — numbered canon, contracted durable sessions, gated waves, adversarial verification, and the deployed system as sole evidence — into an arrangement where the productive abundance of language-model agents serves one intention instead of eroding it. The case suggests that a single designer can govern a production system in their own language, and that as the cost of making collapses, designing the structures of authority becomes a central part of creative work rather than an administrative afterthought. Further cases, external instrumentation, and trials outside the cultural sector are the natural continuations.

acknowledgements

The roles, personas and procedures the sessions rely on are built on BMAD-Method, an open framework for agent-directed agile development, which this project specialises and adapts. The method described here owes that framework its starting grammar; its authors and community are credited with gratitude.

references

• Austin, J. L. (1962). How to Do Things with Words. Oxford: Clarendon Press.
• Cross, N. (2006). Designerly Ways of Knowing. London: Springer.
• Popper, K. R. (1959). The Logic of Scientific Discovery. London: Hutchinson.
• Rittel, H. W. J., & Webber, M. M. (1973). Dilemmas in a General Theory of Planning. Policy Sciences, 4(2), 155–169.
• Schön, D. A. (1983). The Reflective Practitioner. New York: Basic Books.
• Sennett, R. (2008). The Craftsman. New Haven: Yale University Press.
• Suchman, L. A. (1987). Plans and Situated Actions. Cambridge: Cambridge University Press.