Read together, the five essays keep arriving at the same handful of conclusions from different directions. This page states them once, plainly, along with what remains unsolved. Each thesis is the collection's own synthesis — argued from sourced material, but a position, not a citation.
§1The seven theses
- The bottleneck moved.
From writing code to communicating intent upstream and verifying output downstream. Generation is the commodity in the middle — and the measured constraint sits on the verification side.
- Spec renaissance, by necessity.
What was optional discipline for humans becomes mandatory infrastructure for machine collaborators: precise acceptance criteria, test-first, living documentation. The story's "why" survives; its deliberately incomplete card does not — the conversation must leave a written residue. Corrected mid-2026: frontier agents ask clarifying questions, and over-specification is a named antipattern; the target is sufficient precision, not maximal. The upstream shift itself stands.
- Discipline relocates; it does not disappear.
Out of code style, into scaffolding: constitutions and AGENTS.md files, tests as acceptance gates, linters with agent-readable remediation, CI, observability legible to agents.
- Verification is the scarce resource.
Treat AI output as untrusted until verified; review it like a junior's code. Human review of plans and diffs is the guardrail every serious document retains — and the queue where the whole pipeline now backs up.
- Agile values survive; agile mechanisms get rewritten.
TDD becomes the agent's instruction set, pairing becomes plan review, and "working software over documentation" dissolves when the documentation is executable. The waterfall objection is answered by granularity: story-sized micro-cycles, plans as disposable hypotheses.
- Vibe coding is legitimate but bounded.
Prototypes, spikes, UI polish. Production work wants structure. Healthy teams oscillate between the ends of the spectrum on purpose.
- AI amplifies what is already there.
Bad inputs produce faster junk. The win is the 10x organization — scaffolding that lifts everyone — not the 10x developer.
§2Corrections that recurred
Two corrections showed up independently across multiple essays, which is itself a finding. The "zero ambiguity tolerance" claim about agents did not survive 2026 — frontier models ask, and exhaustive specification is now the antipattern arXiv 2603.26233; Thoughtworks Radar, 2026. And the vendor ROI numbers never firmed up: no controlled study isolating spec-driven development's effect on delivery outcomes existed as of July 2026 — essay 04's refrain, adoption verified but value unmeasured, holds across the whole record. A collection that updates its claims in public seemed like the right place to say both out loud.
§3Open problems
| Problem | Status, July 2026 |
|---|---|
| Spec drift — no reverse sync from code edits back to specs | partially matured — manual drift detection
exists (/speckit.converge, OpenSpec); spec evolution unsolved, nothing
self-heals spec-kit repo, 2026 |
| Brownfield codebases exceed context windows; SDD shows diminishing returns there | still open |
| The junior-developer training pipeline — who learns judgment when agents do the routine work? | now measurable — employment for ages 22–25 down ~20% since 2022 Stanford AI Index 2026; the question itself unanswered |
| AI-generated spec bureaucracy — verbose specs nobody reads, the old failure mode with new tooling | still open — the discourse's broader name for the worry is "cognitive debt" Radar Vol 34, Apr 2026 |
| No independent measurement of SDD's effect on delivery outcomes | still open — DORA 2025 measures AI-assisted development generally, not SDD DORA 2025 |
§4Sources
- The theses synthesize the sourced material of essays 01–05; each essay carries its own source block. Figures on this page: DORA 2025; Stanford AI Index 2026 via findskill.ai; spec-kit; Thoughtworks Radar Vol 34 (Apr 15, 2026); arXiv 2603.26233.
- Full method and provenance rules — About.