jaiskills

My agent skills for Claude Code, shipped as one plugin: agents checking agents, verification-first orchestration, model-aware prompting, and OpenAI tooling from the terminal.

These skills exist to fix failure modes I kept hitting with coding agents. Each one is small, composable, and earns its place by closing a specific gap.

Install (30 seconds)

Two ways in, two philosophies — pick one (installing both duplicates every skill).

Claude Code plugin — a managed bundle that updates when I ship, so you subscribe rather than fork:

/plugin marketplace add JairoTorregrosa/jaiskills
/plugin install jaiskills@jaiskills

skills.sh — copies editable skill files into your project (works with Claude Code, Codex, and other agents), so you can hack on them and make them your own:

npx skills@latest add JairoTorregrosa/jaiskills

Or load locally for development:

claude --plugin-dir /path/to/jaiskills

Migrating from ≤0.4.x? This repo used to publish four separate plugins (insistir, metaprompt, constatar, remoto). They are now one plugin, jaiskills. Uninstall the old ones and install jaiskills@jaiskills; every skill and command came along.

Why these skills exist

#1: The agent says it's done — and it isn't

Self-reported success is the most expensive lie in agentic coding.

• insistir (/insistir <task>) — multi-agent orchestration where every worker's output is cross-reviewed by a different agent in an adversarial loop. Tasks cannot be marked complete without a reviewer's APPROVED verdict (enforced by hook). The name is the point: fix, review, judge — insist until verified.
• constatar-plan / constatar-verify (/constatar-run, /constatar-verify, /constatar-audit) — verification-first orchestration through the constatar Rust engine: plans with a 6-rung evidence ladder, grounded verdicts, resumable journals.

#2: One model grades its own homework

Same-model review inherits the same blind spots.

• codex-judge — a cross-provider judge (Codex/GPT-5) scores each review verdict on a weighted rubric and gates auto-approve/auto-revise. The judge never sees implementer reasoning — only diff and verdict — preventing anchoring. Degrades gracefully when Codex is unavailable.
• /advisor — an independent second opinion from a different-provider model on your plan, diff, or question.

#3: The loop plateaus, or worse, games the test

• goal-loop (/goal <goal>) — loop engineering as gradient descent: an agent factory generates goal-specialized agents, then forward → loss → textual gradient → update, with momentum and early stopping. Evidence is split into visible checks (the implementer's target) and held-out checks the implementer never sees; a positive gap between them is treated as reward hacking and yields NOT MET.

#4: Lessons evaporate between sessions

• compound-knowledge (/compound) — captures solved problems as searchable docs in docs/solutions/; a learnings-researcher agent feeds them into future planning.
• file-todos (/triage, /resolve-todos) — review findings become markdown files with a file-name-driven lifecycle (pending → ready → complete), triaged one by one, then fixed by parallel workers.

#5: The prompt wasn't built for the model that runs it

• metaprompt (/metaprompt) — takes a goal, a target model, and a target harness, and produces a complete prompt engineered for that combination, from researched per-model/per-harness guides (Claude, GPT-5.x/Codex; Claude Code, Codex CLI, pi, Amp, and more).

#6: Your laptop is the bottleneck

• remote-agents (/remoto-run, /remoto-status) — orchestrate headless claude -p / codex exec workers on a remote SSH host. File-based job state, jobs survive disconnects, results collected and cross-reviewed across providers. Requires SSH key auth to the host and claude/codex logged in there.

#7: You want OpenAI's models from the terminal — or an image turned into a real page

• askcodex — use GPT-5.x and image models from the CLI with the askcodex binary: one-shot text, image create/edit, models, quota. No API key; it reuses codex login credentials. (Canonical copy lives in the askcodex repo; this one tracks it.)
• image-to-frontend — brief → 4 visual variants → build spec → real React/HTML page, iterated to pixel-close. Image generation runs through askcodex.

Skills

Agents (insistir-worker, insistir-reviewer, insistir-researcher, insistir-learnings-researcher), hooks (completion gate, reviewer Bash whitelist), the bundled Codex MCP config, and the remoto.sh / insistir.py scripts ship in the same plugin. Full inventory in docs/skills.md.

How insistir works

      ┌──────────────────────────────────────────────┐
      │         TECH LEAD (delegate mode only)        │
      │   Coordinates, delegates, synthesizes.        │
      │   NEVER implements, edits files, or builds.   │
      └───┬──────────────┬──────────────┬────────────┘
          │              │              │
   ┌──────▼──┐    ┌──────▼──┐    ┌──────▼──┐
   │Worker A │    │Worker B │    │Worker C │    IMPLEMENT
   └────┬────┘    └────┬────┘    └────┬────┘
        │              │              │
   ┌────▼────┐    ┌────▼────┐    ┌────▼────┐
   │Review B │    │Review C │    │Review A │    CROSS-REVIEW
   │→ REVISE │    │→APPROVE │    │→ REVISE │    (read-only reviewers)
   └────┬────┘    └─────────┘    └────┬────┘
        │                             │
   ┌────▼─────────────────────────────▼────┐
   │        CODEX JUDGE (optional)         │    CROSS-PROVIDER VERDICT
   └────┬─────────────────────────────┬────┘
        │                             │
   ┌────▼────┐                   ┌────▼────┐
   │Fixer A  │                   │Fixer C  │    FIX (fresh agents)
   └────┬────┘                   └────┬────┘
   ┌────▼────┐                   ┌────▼────┐
   │Review B'│                   │Review A'│    RE-REVIEW → APPROVE
   └─────────┘                   └─────────┘

Fresh agents per phase, read-only reviewers, iterative convergence until APPROVED or budget exhausted.

Requirements

• Claude Code 1.0.33+
• Python 3.10+ (hook scripts)
• Optional: OpenAI Codex CLI — enables /advisor, the cross-provider judge, and the askcodex-backed skills
• Optional: the constatar engine — for the constatar skills

License

MIT