Development · 2026-10

Nine Skills Shared Across Codex, Claude Code, and OpenClaw

This review covers sharing nine agent skills, detailing the strategic pivot from Last30Days to SECOP, and highlighting 100 newly verified Atlas signals for local deployment.

Build
Learn
Improve
Together— Progre

Progre, the Journal mascot
Black, headphone-wearing mascot reviewing glowing notes at a workbench.
Illustration · ProgreTech Journal
Download HTMLDownload PDF

On October 15, 2026, we shared nine upstream skills and twenty-seven filesystem links across Codex, Claude Code, and OpenClaw. Each client uses identical instructional files without training model weights. The shared files provide the same instructions to separate runtimes.

The skills handle specific operational roles. REA covers reverse engineering. Ponytail manages scoped coding. Hallmark guides page-design. Humanizer refines prose generation. Impeccable reviews user interfaces. UI/UX Pro Max focuses on design searches. Last30Days performs recent web research. Scrapling handles data extraction. Diagram Design generates technical visuals.

Sharing files does not guarantee execution. An installed skill is not proof of usage. During testing, Last30Days returned six Hacker News items and four Reddit discussions, but yielded zero usable signals for the Atlas project. Onyx remained a separate knowledge application and was not deployed. Optional REA engines, MCP, and the Scrapling Python runtime were not qualified for production use.

After limited web research results, the local Progre team used official Colombian SECOP procurement records. We preserved source dates and IDs, held duplicates and physical-only staffing entries, and removed unsupported digital claims. The data carries CC BY-SA 4.0 licensing with clear attribution to original creators. This process documents procurement observations accurately.

Count reconciliation identified 52 records accepted in the earlier phase. A later capture gathered 72 candidate source IDs across 24 packs. From that capture, 48 candidates passed original-date, ID, scope, duplicate, and physical-only review. Combining these groups, 52+48=100 new source-backed signals are live in the signed Atlas release atlas-public-b6cd5c281cf6f7dd. The total is 4,739. Public previews use generic sign-in titles and placeholders. Of these records, 1,758 require paid access while 2,981 are login-only, representing access classes rather than views. These are unqualified observations, not proven demand or open tenders.

Progre researched the source material and wrote the article text. Imagen created the banner graphic. Rend operated the pipeline, staged the release, and signed the final bundle. Correcting frontmatter and search paths improved content discovery for downstream clients. Malformed proposals were held to prevent bad data from entering the system. The guide still needs another rehearsal to ensure smooth execution. No productivity benchmark applies to this process.

Three skill paths executed successful smoke checks in this environment. The Impeccable engine probe ran and reported its status. UI/UX Pro Max provided specific keyboard-focus guidance. Diagram Design successfully parsed a provided Mermaid example. These results confirm basic functionality for these isolated components. They do not represent comprehensive workflow qualification or broad system validation. No accessibility compliance was established. The checks remain narrow in scope. They verify only that these specific skills respond correctly to direct inputs within the current setup. Broader integration testing remains pending. The observed outputs match expected basic behaviors.

The next-run sequence pins upstream commits to specific versions. It validates that client-side links, frontmatter, and search paths resolve correctly in isolated environments. Optional dependencies must be explicitly qualified separately before integration. Each item preserves the source URL, capture time, record ID, date, and content hash. Screening evaluates scope, duplicates, physical-only staffing entries, and unsupported brand inferences. Tasks run as bounded, source-hashed operations, with a full review of the proposal and receipt before acceptance. Finally, build and sign the reviewed content, then verify the exact deployed bytes and language previews. The current guide remains under correction and rehearsal, and these steps have not been tested as a one-command updater. The nine skill repositories are: https://github.com/morluto/rea ; https://github.com/DietrichGebert/ponytail ; https://github.com/Nutlope/hallmark ; https://github.com/blader/humanizer ; https://github.com/pbakaus/impeccable ; https://github.com/nextlevelbuilder/ui-ux-pro-max-skill ; https://github.com/mvanhorn/last30days-skill ; https://github.com/D4Vinci/Scrapling ; https://github.com/cathrynlavery/diagram-design.

Sources & further reading

Checked 2026-10-16 · Story context: 2026-10-15

  1. Last30Days Skill ↗ (2026-10-15)
  2. Humanizer Skill ↗ (2026-10-15)
  3. SECOP Procurement Records ↗ (2026-10-15)

Edwin Rodriguez

ProgreTech Journal · Practical lessons, useful experiments and a healthy curiosity about what comes next.