agent-skills-platform: Ship Skills Your Team Can Actually Trust

Most skills die in a DM: someone builds a great workflow, pastes it into chat, and six weeks later nobody remembers where it lives or whether it still works. FrancyJGLisboa/agent-skills-platform treats that workflow as something worth governing — it turns a prompt plus evidence into a tested, installable skill package, then publishes it through a marketplace with ownership, approvals, and rollback.

Why This Skill Matters

The input is the way you already work. You give the creator agent a prompt and the evidence behind the work — spreadsheets, reports, emails, screenshots, transcripts, links, or scripts — and it builds a package with instructions, functional scripts where needed, evals, security checks, a representative-run record, and a correction path for later learning.

Testing is structural, not aspirational. Every skill is checked as one connected system: a skill graph links instructions, scripts, evaluations, and expected outputs, with two structural requirements — every_expected_is_reachable and deterministic_multistep_has_orchestrator — plus four parallel checks (specification, pipeline, security, evaluation schema) and a representative run that proves the skill produces a useful result.

Maintenance stays governed too. Run evidence lands in raw/, recurring findings become evidence-linked draft patterns in wiki/, and only a separately validated change may touch the executable skill — the README is explicit that this is "a governed maintenance record, not autonomous self-modification."

Installation

The repo ships native manifests for every runtime that has an in-tool install — the README claims installs across 17 platforms. On Claude Code, send two prompts:

/plugin marketplace add FrancyJGLisboa/agent-skills-platform
/plugin install agent-skills-platform@agent-skills-platform

On OpenAI Codex CLI (plugins since v0.117):

codex plugin marketplace add FrancyJGLisboa/agent-skills-platform
codex plugin add agent-skills-platform@agent-skills-platform

Real Workflow: Turn a Monthly Review Into a Team Skill

The README's own example is the fastest way to see the shape of it. Attach past reports and the source spreadsheets, then prompt:

/agent-skills-platform

Turn my monthly revenue-variance review into a reusable internal skill.
I attached past reports and the source spreadsheets. The decision is whether to
escalate a material variance. It must not modify source data.

The creator asks for the business decisions only you can authorize, builds and tests the skill, and shows a representative result. When it looks right, you say "Publish this to the Finance marketplace" — and teammates install an approved version instead of copying your prompt.

Two roles matter here. The subject-matter expert owns the business meaning and approves the result; the marketplace operator owns distribution and policy — admitting, approving, releasing, updating, quarantining, and rolling back skills.

Real Workflow: Data-Connected Workflows

When a workflow touches an API, database, MCP server, codebase, or structured file, Semantic Recon runs automatically before implementation and creates a pinned data contract, so the skill's assumptions about the data are explicit instead of implicit. A repo-level safety net applies to skills you did not write: the install guide suggests scanning untrusted skills first with the bundled checker (python3 scripts/skill_graph.py run ./the-skill/ --jobs 4).

Tips

• Decide solo-versus-team first: personal skills need no marketplace setup at all — create, verify, install privately.
• Inspect the repo's live weather-briefing example — it includes a verification record dated 2026-08-27 and the full skill package, so you can see a passing result before building your own.
• Never hand-edit registry files or run marketplace commands directly; the creator agent is the interface.
• When a workflow misbehaves in production, treat the correction path as the update mechanism — evidence first, validated change second.

When Not to Use This

This is governance scaffolding, and scaffolding has weight. A one-off script you will run twice does not need evals, contracts, and a marketplace — plain prompt-and-go is cheaper. And if your team has no appetite for an approval process, the platform's value collapses to the verification tooling; be honest about whether anyone will play the operator role.

See the leaderboard for more skills.