Dididecks vs Instadeck.
One product, two approaches.

READMEs on both sides route the reader elsewhere quickly: hers to product/, with a release-safety rule; ours to the specs, the splash, the client sites and the skills snapshot. Ours has drifted: it calls a mock-backed shell “active”.

Agent scratch kept out of git. Both ignore the coding agent’s local folder. Hers also ignores /docs and /scripts by default, force-adding only the records she wants tracked, and labels Bitácora logs not product or agent context: an explicit line between notes and record that ours doesn’t draw.

Deploy config. Hers carries policy in the config (API startup must not race migration); ours is small because static sites need little.

Agent Skills as domain know-how. Both use SKILL.md for craft rather than plumbing: hers for VC analysis and slide craft (ic-critique, why-now), ours for brand, design contract, ingest and imagery. Hers are short (271 lines for 19) and imperative; ours are long and explain why.

Chromium as the judge of a slide. Hers proves every render before publication and measures geometry and asset visibility. Ours drives a running site to PNG, SVG, PDF, PPTX and Keynote, and the Play-UI rule (No responsive CSS. No JS in the slide itself) exists so export is faithful. Hers uses the browser to check; ours uses it to produce.

Instruction files. Both built three levels: hers root, backend/ and frontend/; ours anchor, ai-labs and dididecks-ai. Hers contradicted itself, so she removed it from main and wrote a shorter reset pinned by a test. Ours contradicts itself differently, with stale skill paths. Ours orients a cold agent in one read, including why; hers can’t silently drift.

Specs. Her spec of record has a precedence order for authority, change control, a conformance ledger rating each requirement Implemented, Partial or Not proven, and a definition of done. Our main spec is a north-star document, still Draft. Ours keeps questions open while the product is being discovered across engagements; hers closes them so an agent can’t reopen them by accident. Hers costs earlier reasoning, which survives only on archive branches; ours costs an agent knowing what has been decided.

Decision records. Her dated records open by saying what they do not authorize: not approved for merge, not deployment authorization. Ours record reasoning and loose ends. Her PR 39 recovery note is a candid account of a reversal.

Living maps, runbooks, session memory. Our sitemap has a file per route with typed frontmatter; her runtime inventory says it does not claim browser acceptance. Our browser-drive blueprint is tree-wide; her live runbook is stricter: Do not call the release complete unless the command exits zero. Her Bitácora stays close to the operator and out of every agent’s context; our changelog is shared, searchable, and read by the deck agent.

Briefs for a builder, a mirror image. We wrote data models so she “stops guessing at shapes”; she wrote a Codex pack to brief Codex on the dididecks model. These documents were the real interface between the two silos. Git: hers makes a change reviewable as a unit; ours makes the reasoning readable in git log. Coding agent: Codex and Claude Code; neither side’s artifacts are tied to its tool.

The Markdown document chain, the strongest convergence. Hers: 01-source.md through 06-DESIGN.md, then slides/sNN.md, each creative source plus typed front matter. Ours: narrative files with typed frontmatter and a 592-line DESIGN.md for one client. Hers adds earlier stages as named files, hashes and evidence IDs; ours keeps files across weeks of iteration and many variants. Her spec names a principle we never wrote down: Markdown is not factual authority … Strong typing is required at safety/execution boundaries.

The generator. Hers is a service that runs unattended for any customer at any hour, in a large file that is hard to load. Ours can do anything Astro can render, and needs an operator for every change. The contract: her canonical prompt is fixed per release and customer-safe (default intent: Make this deck much better.); ours is a method an operator adapts per client, and it explains its reasons.

Spec row to component. In June our exploration proposed a structured spec row compiled deterministically into a component. In August her prompt-system decision proposed a DeckSpec plus a deterministic compiler and left it unmounted. Both reached the same hybrid separately; neither has made it the main path.

Judgement, knowledge, diligence. Her visual audit scored slides on eight dimensions, once; our rank pill records a four-state verdict per slot, continuously. Her knowledge is broad and pre-written (195 files); ours is narrower and grows from engagements. Her Due Diligence findings stay off the investor slides; our citation popover keeps sources one click away on the slide.

Repo memory as product memory, refused, the deliberate counter-choice of the study. Hers makes easier: customer runs can’t be steered by operator notes; memory is per customer and typed; a new agent can’t mistake history for authority. Harder: development reasoning lives in PR titles, dated docs and a local log, so a cold agent has less to read. Ours makes easier: one place for both kinds of memory, cheap to add, readable by any agent, compounding across clients. Harder: an agent can treat an old exploration as current, and nothing separates a client’s confidential notes from the method.

Hers: release paperwork for migrations and four services against customer data; patch transport through CI as a workaround for her agent environment; pivots kept as archive branches. Ours: pseudomonorepo rules, the shared skills library, the corpus and publishing, which exist because of seven client repos inside a 27-project tree.

Where the generator’s memory lives. Her knowledge loader excludes operator notes from every product knowledge lane; our generator reads the client folder, because it is the operator’s agent. Where generation will run, the deepest strategic difference: our spec bets on a local-first native app with embedded Claude chat, because the paying user won’t open a terminal and confidential material stays local; hers bets on a hosted web app. Ours keeps the agent’s flexibility and moves it closer to the user; hers keeps the user’s simplicity and constrains the agent.

CI that runs tests, and tests at all. The shell has shipped at least four versions to seven consumers with typecheck as the only automated check; her one workflow runs about 27 test files with no paid calls. Tests that pin documents: hers assert that the spec and genesis still say what they must, and that the spec’s prompt equals the runtime prompt; we rely on discipline and frontmatter. A review packet, headed unfinished implementation. Do not merge or deploy this snapshot: low priority, but it matters when someone else picks up unfinished work.

Drift noted in dididecks-ai (not fixed here): CLAUDE.md lists two submodules where there are seven client sites and a corpus, and an old skills path; a README claim about a mock; architecture links to files that don’t exist; five skills in the snapshot differ from the anchor, and one isn’t linked anywhere a session can load it; a filemap last regenerated 2026-06-07.

On our side, the main lesson of this layer. A frozen, provider-free corpus of past passes and failure shapes, where we have 51 exported slides and nothing that checks a shell change against them. Acceptance runners with 11 conditions, where ours is a human walk-through. Measured failure, then a bounded repair, at exactly the Play-UI step CLAUDE.md calls hardest. Tests that keep the runtime prompt equal to the spec, where our equivalent, the skill pack, has drifted.

On her side, a flow from ours. Variants side by side. Her genesis names the problem: Sometimes the LLM produced an excellent redesign. Sometimes it produced generic layouts. Her answer is constraint and measurement around one result per run; ours is optionality, every slot with several variants ranked by a human. Since her §21.2 acceptance is a human judgement, two or three treatments for the weakest slides would give it something to choose between. The cost is paid model calls, which her spec guards.