skill-harness

Handle /loop-style requests for finding, adapting, drafting, and preparing bounded agent loops without treating external loop catalogs as authorization.

Turn external skill, rule, plugin, MCP, task-memory, and multi-agent workspace patterns into safe fixture coverage rather than live dependencies.

Check whether host instruction surfaces such as AGENTS.md, CLAUDE.md, Codex agents, Cursor rules, Copilot instructions, and SKILL.md files have drifted apart.

Review skills generated from recorded workflows before they are installed, shared, or treated as reusable automation.

Review skills, agents, rules, plugins, and MCP/tool packs for license, source, script, permission, and supply-chain risk before adoption.

Choose the right task-memory and issue-state pattern for a repo: Beads, lightweight files, external tracker integration, or no durable task memory yet.

Review third-party skill, agent, rule, and plugin repos for format fit, provenance, and first-party adoption paths without treating them as direct harness dependencies.

Synthetic skill for intake scanner coverage.

Synthetic risky skill for intake scanner coverage.

Produce safe, self-contained HTML review artifacts from canonical project sources.

Shape product, business, data, research, UX, and mockup artifacts as agent-readable sources with generated visual human review surfaces.

Align design artifacts with code tokens and component APIs so spacing, color, type, and states stay consistent through implementation.

Translate a Figma screen or component into an implementation plan covering structure, states, constraints, and component boundaries.

Choose the right developer artifact type, source of truth, and review surface for a task.

Shape source-backed model and diagram review artifacts for architecture, UML-style, C4, dependency, and workflow views.

Review agent memory, handoff, retrieval, and learning designs for usefulness, provenance, privacy, staleness, and poisoning risk.

Review agent run traces, eval summaries, tool logs, and handoff evidence before accepting a workflow result or self-improvement proposal.

Shape ambiguous or frontier-agent work into a bounded agent task with outcome, scope, context, tools, autonomy level, and verification evidence.

Review an agent workflow for where it may act autonomously, where it must ask, and where human approval is required.

Plan the context, artifacts, memory, retrieval, and compression surfaces needed for an agent workflow or long-horizon task.

Review or design multi-agent workflows with clear ownership, handoffs, evidence gates, conflict handling, and minimal coordination overhead.

Design governed self-improving agent loops that sense work, model failure modes, act in small slices, gate evidence, and turn learnings into tracked proposals.

Design least-privilege tool access for agent workflows, MCP servers, connectors, browser automation, shell commands, and external side effects.

Assemble approved demo media, canonical source specs, evidence links, provenance, and promotion notes into a release or handoff bundle. Use when Codex needs to prepare demo-machine, manual QA, or generated media outputs for docs, launch pages, changelogs, social previews, or another agent.

Create safe static review surfaces for demo and QA media. Use when Codex needs an HTML or Markdown review page comparing demo videos, source specs, QA reports, screenshots, storyboards, quality findings, and evidence links from demo-machine, manual-qa-machine, Playwright, or video analysis outputs.

Plan no-caption slideshow-style MP4s and frame reels from demo-machine screenshots, storyboard frames, selected video spans, or manual QA evidence. Use when Codex needs a polished still-frame walkthrough, hero reel, or reviewable visual summary without recapturing the browser flow.

Plan short silent demo cuts, slideshow edits, hero loops, and social clips from existing demo-machine or QA evidence. Use when Codex needs no-caption polished videos, frame-based reels, poster frames, or audience-specific excerpts derived from `events.json`, screenshots, storyboard evidence, or rendered demo runs.

Package completed demo-machine runs, browser capture artifacts, or evidence-backed product demos into source-linked handoff bundles. Use when Codex needs to organize `.demo.yaml`, `events.json`, `quality.json`, screenshots, storyboard outputs, rendered videos, review prompts, or release-ready demo assets without losing provenance.

Convert manual QA flows, QA reports, screenshots, network/console evidence, accessibility findings, or bug reproduction steps into demo-machine specs, reproducible demo plans, or short evidence-backed repro clips. Use when Codex needs to turn manual-qa-machine output into a product demo, release proof, or issue reproduction asset.

Verify that developer artifacts are grounded, current, and safe to hand off.

Assemble the minimal artifact bundle needed for another agent or human to continue work.

Inspect failing GitHub Actions checks, isolate the actionable failure, and turn it into a concrete fix path with verification steps.

Address GitHub PR review threads or issue comments with explicit comment selection, repo-grounded fixes, and concise reply-ready summaries.

Run implementation work from a tracked issue with explicit scope, branch intent, validation steps, and closeout evidence.

Check whether a change is ready for review by testing the claimed behavior, summarizing risk, and naming any gaps reviewers should know about.

Decide how an external app or service should be integrated, including boundaries, auth scopes, ownership, and failure handling.

Plan an MCP server with clear tool scope, auth model, error behavior, evaluation approach, and rollout boundaries.

Review third-party skill repos for format fit, provenance, and first-party adoption paths without treating them as direct harness dependencies.

Run the noslop pre-commit quality gate, interpret failures, and enforce content-aware protection rules before every commit.

Run the noslop pre-PR quality gate and handle the noslop-approved escape hatch for intentional config weakening before opening a pull request.