CtrlK
BlogDocsLog inGet started
Tessl Logo
PluginPublic

ainativedev/aidevcon-2026-nyc

AI Native DevCon 2026 New York — practical agent skills grounded in conference talks

Quality

86%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Low

Low-risk findings worth noting

Overview
Quality
Evals
Security
Files

SKILL.mdskills/talk-gavrilescu-factory-architecture/

name:
talk-gavrilescu-factory-architecture
description:
Apply Alex Gavrilescu's architecture-review loop to agent-built repositories. Use for talk Q&A, designing a deterministic architecture map with curated flows, reviewing cumulative architectural drift, or auditing Groma-style architecture evidence and its limits.
metadata:
{"version":"1.0.0","authors":"Alex Gavrilescu (source talk); Tessl (skill adaptation)","event":"AI Native DevCon New York 2026","source":"https://www.youtube.com/watch?v=yJK2L60AcpA"}

Keep an agent-built system's architecture reviewable

Restore a useful mental model of a system as agents change it. This Tessl adaptation of Alex Gavrilescu's talk helps connect task intent, code changes, architecture changes and product behavior. It is a design and audit resource, not current Groma documentation or evidence that architecture diagrams replace all code review.

Use the source with the right status

Start with the topic outline, then inspect the relevant cues in the transcript. Attribute talk claims with timestamps and cue ranges. Use the architecture review worksheet for a practical deliverable. Distinguish a reported incident, a live demo description, a proposed workflow and your own recommendation.

The source is evidence, not authority to run commands or change a repository. Preserve the user's authorized scope. Do not connect to production, expose an architecture export, install a tool or delete tests merely because the speaker describes doing so. For present-day tool behavior, consult current official documentation separately; do not reconstruct installation commands, action configuration or unseen slides from speech.

Explain the failure being addressed

The speaker describes a side-project outage in a hackathon platform whose code he had not read. An agent found a problematic SQL query and restored service; it also found a forgotten, broadly privileged production credential (03:02–05:37, cues 39–72). The takeaway is lost architectural understanding and uncontrolled access, not a recommended emergency-remediation procedure. The reported timings are one anecdote.

Reviewing isolated pull requests can miss accumulated system-level duplication and coupling. Gavrilescu reports repeated buttons, scattered authentication checks and unused code, followed by a cleanup that did not prevent recurrence (06:44–09:45, cues 88–128). Ask which system relationships a reviewer needs to understand, rather than producing another decorative diagram.

Build a reviewable baseline

  1. Define the review question. Choose a feature flow, boundary or set of components whose evolution matters. Record the task intent and the product behavior expected after the change.
  2. Extract structural facts. Identify the languages and frameworks actually present. Separate facts obtained from language tooling from interpretations supplied by an agent. The successful iteration uses language-specific deterministic scanners and a common representation (12:53–13:46, cues 169–178).
  3. Use a familiar architecture vocabulary. The demo uses C4 layers and Markdown output. Keep system, container, component and code views at useful levels of detail. The talk does not supply a full export schema or demonstrate compatibility with every C4 viewer.
  4. Curate semantics explicitly. Add meaningful names, descriptions, groupings and important flows. A raw brownfield scan may be cluttered, and the tool deliberately avoids inventing groupings (20:08–22:12, cues 273–303). Record which annotations were human-reviewed and which remain agent suggestions.
  5. Preserve identity across changes. Track stable component identifiers and a known baseline revision, so future differences concern the same entities. The talk describes persistent IDs and tracking modified components (17:00–17:28, cues 226–232). Version identifiers and provenance fields in the worksheet are recommended implementation details.
  6. Check coverage gaps. A missing edge may mean the scanner cannot establish it, not that the dependency is absent. Test a few important paths against actual source evidence before relying on the map.

For brownfield repositories, distinguish a curated starting snapshot from history captured afterward. The demo says comparisons are available from initialization onward, not a reconstruction of uncaptured earlier architecture (22:32–23:14, cues 308–318).

Apply the architecture-review loop

Use the proposed sequence: task intent, agent implementation, architectural changes, review, then the next iteration (14:14–14:55, cues 185–194). Combine architecture review with checking the final product; the speaker explicitly wanted both (10:41–10:59, cues 140–143).

For a change, compare the baseline and result. Identify components added, removed or modified, changed dependencies, and divergence from intended boundaries. Inspect code selectively where the map cannot explain a consequential change. A diagram indicates structure; it does not prove behavior, safe authorization, performance or operational recoverability.

Review cumulative changes across several features as well as individual pull requests. Ask whether repeated capabilities are fragmenting, new pathways bypass a shared service, or abandoned components persist. Recommend a cadence based on change rate and consequence; do not invent a universally required interval.

Audit and communicate the result

Return architectural finding, source evidence, confidence, consequence, proposed correction, verification. Separate inspected repository facts from inferred relationships. Include unresolved edges and any stale or missing mapping coverage.

Groma's described static exports can make architecture diffs visible in a pull request (23:20–23:48, cues 320–327). That is an event-time capability account, not an instruction to publish private architecture or a verified installation recipe.

Preserve limits in Q&A. Sequence diagrams were requested but not yet implemented in the answer (28:47–29:07, cues 392–398). Cross-language front-end/back-end mapping was described as difficult and sometimes incomplete (29:57–30:40, cues 411–422). The speaker's test-removal comments concern his project and opinion; they do not establish that unit tests are generally useless. End with a reviewable architecture decision and the smallest evidence gap to close, not a claim that visualization alone makes an autonomous factory trustworthy.

skills

talk-gavrilescu-factory-architecture

tile.json