Build1 distinct publisher3 min readUpdated
A developer reports driving a CASE repository of 24,483 catalogued elements from sqlplus alone: 722 CI_* views to read, about 200 CIO* packages to write, and no surviving documentation.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
A developer working on a large insurance system published an account on dev.to of an AI agent that designed an entity, a table, a sequence, two triggers and a complete Oracle Forms definition inside Oracle Designer 10g, moved a table between applications, and debugged the generated form to a working runtime, without opening the Designer GUI once [5]. Designer is a CASE tool from the early 2000s whose support ended in 2013 [1], which puts it roughly twelve years past its last patch [21] and makes it the kind of system usually filed under "replace eventually" rather than "script this week".
The repository in question is not a toy. It holds 2,741 entities, 3,277 table definitions, 12,841 PL/SQL modules and 5,624 forms [2], about 24,483 catalogued elements in total [19], with roughly ten developers working in it daily [3]. Everyone there, the author included, had assumed the tool was its GUI: clicking, dragging, repeating [4].
What the account documents is a three-layer API sitting in the same Oracle schema as the data. Reading is 722 CI_* views, where every entity, column and form item is a row [7]. Writing is about 200 CIO* packages, one per element type, each with ins/upd/del/sel and a typed record [8]. Transactions go through CDAPI: open_activity, then the changes, then validate_activity, which applies Designer's own validation rules, then close_activity, with abort_activity rolling the repository back on failure [9]. That is roughly 922 SQL-callable objects [20]. Forget jr_context.set_workarea and you get CDR-00100 [10].
No documentation for any of this survives, according to the author, and it did not matter: the package specs are readable from all_source, so the API documents itself [11]. The working loop was four steps: read the API from the database, read an existing element as the template, write in small validated transactions that print violations and abort on failure, and verify through the CI views before moving on [12]. The second step is the interesting one for anyone maintaining a system with no written standards. There was no house-standards document, but there were 3,277 existing tables, and naming conventions, audit columns, domains, window sizes and preference sets were all read out of elements colleagues had built over twenty years [13]. That included work normally done by a wizard, the Database Design Transformer, which handles table-to-entity mapping [14], plus a sequence with its ID trigger, audit columns and their trigger with PL/SQL bodies written through the RMOTEXT API, and a table ownership move that left a shortcut behind in the original application [15].
The boundary is where the tooling stops being data. Generating the actual Forms binary needs a client-side tool, so the author pressed Generate; when the generator failed, the agent read its output and the .err file from disk, fixed the causes through the API, and the cycle repeated until the form ran [16]. One failure was visual rather than textual: fields stacked on top of each other because items without display_width fall back to the column width, diagnosed from a screenshot and fixed by setting widths through the API [17]. Around twenty distinct errors came up over the two days, each diagnosable from the message plus the state of the database [18]. Total human GUI time: two buttons [6].
Worth watching: whether the same read-template-write-validate loop holds on repositories where twenty years of convention are inconsistent rather than merely undocumented, and how much of a legacy stack's remaining GUI dependency is, like the Forms generator, a compile step rather than a design step.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
Oracle Designer 10g is a CASE tool from the early 2000s and support for it ended in 2013.
The Designer repository at the author's job holds the design of a large insurance system: 2,741 entities, 3,277 table definitions, 12,841 PL/SQL modules and 5,624 forms.
Everyone at the job, including the author, assumed Designer work was done through its GUI: clicking, dragging, repeating.
Within two days an AI agent with SQL knowledge and database access had designed an entity, a table, a sequence, two triggers and a complete form definition, moved a table between applications, and debugged the generated form to a working runtime, opening the Designer GUI zero times.
Everything the author personally did in a GUI fits in two buttons: Generate, and starting the form.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Specific and partly checkable, but single-source
The post supplies unusually concrete, independently checkable artifacts: named API layers with counts (722 CI_* views, ~200 CIO* packages), a runnable CDAPI write template, a specific error code (CDR-00100), an all_source introspection query, and a named root cause (missing display_width). Anyone with a Designer schema can verify the API surface claims. What cannot be verified from the supplied material is the outcome narrative — the two-day timeline, the zero-GUI claim and the ~20 errors are first-person assertions with no second source, log excerpt or third-party reproduction.
One shop, one experiment, one skill release
Adoption evidence is confined to a single employer where the author works, and even there the ten daily developers are described as GUI users; the agent workflow is the author's two-day exercise. The only outward artifact is a freshly published MIT-licensed Claude Code skill with no download, star, contributor or third-party usage figures in the supplied material.
Mildly overstated headline, disciplined body
The body is notably self-limiting: it concedes the Forms/DDL generators are client-side, that a human pressed Generate repeatedly, that diagrams remain manual, and that roughly 20 errors had to be worked through. The overshoot sits in framing — a headline and skill blurb generalizing a single two-day, single-form exercise into 'develops in it anyway' and applicability across 'many legacy Oracle shops', plus the implicit suggestion that agentic access substitutes for the GUI at production scale rather than in one demonstration chain.
Self-reported success promoting the author's own skill
The author is both the operator of the experiment and the publisher of its output: the post concludes by promoting an MIT-licensed Claude Code skill and handbook derived from the exercise, on a developer platform where such posts build personal reputation. There is no disclosed vendor sponsorship, paid product or commercial relationship in the supplied material, and the open-source MIT licensing limits direct monetary upside, so the incentive is reputational and distribution-driven rather than financial.
Low — one self-published account
Confidence is capped by structure: a single publisher, a single author, no corroborating source, and outcome claims that only the author could observe. The structural API claims are highly plausible and testable, which keeps confidence off the floor, but nothing in the supplied material independently validates the productivity or reliability of the agent workflow.
build
Thirteen tasks green, then "give up (Recommended)" on the one that needed understanding1 distinct publisher
leadership
A Government Switched Off Two Frontier Models. Your Board Will Want The Fallback Plan.1 distinct publisher
build
Opus 5 absorbed your verify prompts. The reading is still on your desk.1 distinct publisher
build
Ornith-1.0's benchmarks are fine. Ollama can't parse its tool calls.1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · August 21, 2026