Context Layer

I created Context Layer because working with AI kept making project continuity my job. A conversation could produce useful work, but decisions, unfinished plans, evidence and the reasons behind a change needed to survive the conversation. I wanted to resume a project without reconstructing it from chat history or trusting the agent’s latest summary. My answer began as ordinary project documentation stored with the code. Over time, I turned that practice into a scaffold, then a set of workflows, and finally a plugin.

The earliest concrete ancestor recovered for this history is an October 25, 2025 Family Shapes commit. It organized numbered project documents into active, completed, backlog and archived folders, and connected those records to the roadmap and changelog. Permissions, sharing, tree-builder UX and union-node work each had a place. This was the useful idea underneath everything that followed: give a piece of work a stable identity, keep its plan with the repository, and move the record as its state changes. By February 6, 2026, my reusable docs-system-scaffold skill explicitly described itself as modeled on the Family Shapes DOCS structure.

The first reusable scaffold created the documentation structure, a VitePress site and scripts for manifests, links and frontmatter. I used it in the Notion project-dashboard automation repository, where the full build plan became Project 1001 instead of remaining in the conversation. In February I also asked for automatic docs-site updates to become a default of the skill. As I reused the scaffold, it acquired agent instructions, timestamped checkpoints, evidence storage, product documentation and review routines. By the May 12 Git snapshot it was called Product Operating System Scaffold. The broader name reflected how much responsibility the original documentation skill had accumulated.

In March, while working on the Rock Paper Scissors Shopify storefront, I asked for two additional states. In-review meant implementation could be finished while still waiting for me to inspect it. Blocked meant the agent needed a decision, permission or other input before it could proceed. I wanted an agent to continue useful backlog work without pretending I had accepted the previous result. The first lifecycle extension was installed in that repository, validated there, then folded into the standard scaffold. The separate retrofit skill remained for older projects. This is one of the clearest examples of a real project changing the reusable system.

In April I asked for a dashboard over the project records and explicitly checked that Git would remain the source of truth. The dashboard became generated HTML, with the project Markdown underneath it. I also introduced literal MAGGIE TODO callouts for the testing, evidence, approvals and external work that only I could supply. Those follow-ups needed to be visible when an agent handed work back. Alongside this, I tried a separate repo-pm-orchestrator / repo-pm-scaffold-v2 prototype with program pages, dotted project numbering and a sidebar. The conversation records its implementation, but I have not established that this experiment became the direct ancestor of the shipping plugin.

On May 13 I evaluated the large scaffold and asked how to improve reliability without breaking a system I already depended on. The problem was partly invocation: an agent could forget what I meant by “create a review board,” even when the repository contained the workflow. I wanted focused skills that would lead it to the right instructions. I chose the name Context Layer that day. The package separated setup from ongoing actions, moved detail into references, made the project dashboard required, standardized DOCS and the root project template, and codified parent–child projects. The earlier scaffold was preserved as a rollback copy. The initial package grew to ten skills; dashboard backfill made eleven in June. The current package is version 0.1.0.

Maggie Todo was a deliberately smaller adaptation. In May I asked for one place to keep personal and cross-project follow-ups, without creating another hosted application. The implementation used canonical Markdown todo records, a generated local dashboard and the same MAGGIE TODO convention. One recorded batch connected a source conversation, the Shopify project intake, a new image-evidence capability and the screenshot intake that depended on that capability. The dependency and the human follow-up remained attached to the work. Hyphenomenon uses the same foundation to connect project records, evidence and publication checkpoints; this page is being maintained through that system.

A September 24 Frill audit shows the system doing the less glamorous work that matters during a handoff. It reconciled sixteen existing project documents against release evidence, an issue register and the original requests. It found old launch blockers that had already been resolved, distinguished deployed features from historical “local only” notes, and preserved my explicit pause on builder work. The delivery included a request ledger, an action queue, source inventories and a fresh-agent brief. It also exposed duplicate project numbering and stale lifecycle labels that needed correction. Context Layer did not prevent all drift; it gave the audit records to reconcile and a durable place to put the corrected state.

I expected the portfolio to need only small updates, so I did not maintain the full standard layer in the redesign checkout. By October, studies, review surfaces and handoff packages had accumulated into a roundabout process: repeated work, inconsistent storage and one shadow product after another. The October 11 conversation records my realization and the installation that organized the work into five owning projects while preserving 690 existing documents and study files. The recovered history adds a qualification: a lightweight scaffold had been installed in a separate portfolio worktree in May, with its dashboard and other modules turned off. That earlier partial setup did not establish a maintained standard layer in the October redesign checkout.

By May I was already asking whether native Codex features would make my custom system obsolete. In June, while trying Cursor, I called Context Layer the foundation of my process. By October I was still seeing the difference in a gpt-6-astra high session, despite the richer context support around it. My conclusion is practical: I need an inspectable record of what a project is, what was decided, what has actually shipped and what still needs me. Better session context helps an agent work with that record. The July compliance-playground investigation also makes the limit clear: having the files does not guarantee that every model will discover or follow them. The evidence here shows reusable workflows, concrete handoffs and recoverable project state; it does not establish a measured productivity uplift.