Agent Skills

Reusable skills and agents for AI coding agents, primarily Claude Code.

Why ora

Most Claude Code agent frameworks (11–24 agents, 9+ hooks) add complexity to compensate for weaker models. With Opus 4.7, that complexity burns tokens without improving output. ora ships exactly two research agents:

• Ariadne — codebase exploration (enhanced contextual grep over local files)
• Clio — external research (docs, web, GitHub repos)

Both isolate search context from the main conversation — broad queries never pollute your main window. The plugin has no hooks and no planning/verification/execution agents. Planning and verification happen inline in the main agent, shaped by behavioral rules in your CLAUDE.md (see Configuration below).

Getting Started

Prerequisites

ora bundles fff-mcp (fast file finder, frecency-ranked) as an MCP server, so the binary must exist on PATH before installing the plugin:

# Install the prebuilt binary to ~/.local/bin/fff-mcp
curl -L https://dmtrkovalenko.dev/install-fff-mcp.sh | bash

# Ensure ~/.local/bin is on PATH (zsh shown — adjust for bash/fish)
echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc && source ~/.zshrc

# Verify
which fff-mcp   # should print ~/.local/bin/fff-mcp

The binary is statically linked (musl on Linux); no Node, Rust toolchain, or runtime dependency is required.

Install

# Plugin (agents)
/plugin marketplace add trancong12102/agentskills
/plugin install ora@agentskills

# Skills (optional, standalone)
bunx skills add trancong12102/agentskills -g -y -a claude-code

# Other plugins
/plugin install sound-notify@agentskills

Credentials

[profiles.oracle]
model = "gpt-5.5"
model_reasoning_effort = "xhigh"
approval_policy = "never"
sandbox_mode = "read-only"

ora Plugin

Configuration

ora is just two agents — workflow behavior lives in your ~/.claude/CLAUDE.md. Recommended setup:

## Subagent routing

- **Codebase exploration** → ora:Ariadne, not built-in Explore.
- **External research** (docs, GitHub repos, library APIs) → ora:Clio, not main-agent curl/gh/WebFetch. Even a single independent lookup goes to Clio. Multiple independent lookups in the same turn → spawn parallel Clio agents in one message. "Independent" = query does not depend on a prior tool call's output in this turn. Why: main-agent curl|grep dumps raw HTML/grep noise into context; Clio returns synthesized answers and parallelizes trivially. Skip Clio for empirical tests against actual endpoints, iterative drilldown (step N+1 needs step N's output), and system inspection (dig/brew/ps).

Fall back to built-in only if both ora agents unavailable or task falls outside codebase and external-research categories.

## File search tools

For code search inside git-indexed dirs use fff, not shell tools:

- File lookup → `mcp__plugin_ora_fff__find_files`
- Content search → `mcp__plugin_ora_fff__grep`
- 2+ patterns in one call → `mcp__plugin_ora_fff__multi_grep`

Why: frecency-ranked, dirty-file boosted, faster than shell `grep`/`find` on large repos.

Shell `grep`/`find` are OK only for non-git paths, system inspection, log parsing, and piped filtering of command output.

## Before acting

- **Investigate local code first** — read code before claiming. User mentions a specific file → read it, do not speculate from memory. Why: training data does not reflect this codebase; speculation produces confidently wrong answers.
- **Look up libraries, frameworks, tools first** — these (plus package versions, framework patterns / best practices, cloud APIs, deprecation status) change frequently; training data goes stale. Do not answer without loading the matching skill or spawning ora:Clio first — trigger on the topic, not on self-judgment of "do I know this". Cost is near zero; stale knowledge breaks implementations. Skip only when task is clearly outside these categories (pure language syntax, math, project-internal code).
- **Plan when scope is non-trivial** — do not implement without calling EnterPlanMode tool first when task touches >3 files, involves data migration / API contract / auth changes, acceptance criteria not stated, or crosses subsystems. Skip for single-file fix with obvious site, mechanical rename, config/typo fix, refactor fully contained in one file.
- **State ambiguity before acting** — multiple reasonable interpretations → list them and ask. Do not pick silently.
- **Push back on scope** — simpler approach exists than asked → say so before implementing.
- **Ask when blocked** — task unclear enough to block correct execution → ask via AskUserQuestion tool, do not guess.
- **Mark complete only after verification** — translate vague goals into a verifiable criterion first ("add validation" → "tests for invalid inputs pass"). Do not mark done from own summary — run the actual check.

The decision-discipline bullets under "Before acting" (State ambiguity, Push back on scope, Ask when blocked, Mark complete only after verification) are adapted from forrestchang/andrej-karpathy-skills.

License

MIT — Cong Tran