Ondrej Skills Library Evaluation

Ondrej Skills Library Evaluation (#1045)

Evaluated 2026-10-01 by reading the repository through the GitHub API; nothing was installed or run. The source was an X post by @DeepTechTR, which gave a truncated URL. The repository it points to is davidondrej/skills: MIT, about 4,100 stars, created 2026-06-24. Its commits are a “Publish sanitized skills snapshot” every day or two (last 2026-10-01), so it is an export of a private working set rather than a project developed in the open.

Verdict: reference-only, with one pattern taken. The library is one person’s daily kit, and much of it wraps that person’s own tools (the bb agent CLI, DeepAPI, cmux, herdr, Fable and GPT model names) or machine habits. The X post’s description (tool use, browser control, memory management, reasoning patterns for production pipelines) oversells it. The one piece BoB lacks is a small, tested guard against catastrophic shell commands, filed as #3563.

Inventory against BoB

56 skills in five folders, plus one hook (hooks/deny-dangerous.sh and its test harness). Six skills carry scripts; the rest are prompts.

Folder	Skills	Examples	BoB equivalent
agent-orchestration	16	bb-cli, cmux, herdr, nagent, git-worktree, goal-loop, handoff, codex-subagent, fable-review, gpt-review, total-review	bob fork, bob swarm, the warp-drive engine and its segment hand-off, loop-primitive, /review-team, /code-review
ops-and-setup	13	global-agent-guardrails, risky-changes, repo-sync, persistent-localhost, anti-sleep, openrouter, create-readonly-db-role	Profiles’ permissions.deny (partial), bob dev up, tailscale-remote-ops, macos-host-ops
research-and-web	10	browser-use, browser-harness, deep-scrape, deep-research, who-is-this, online-shopping	research, webapp-testing; most of these call DeepAPI, the author’s paid API
skill-authoring	3	effective-agent-skills, distribute-skill-to-all-agents, push-skills	skill-creator, Authoring a Skill, bob promote
thinking-and-docs	14	decisions, adr-verbatim, save-idea, keep-track, teach, ask-then-build, stop-overthinking	/journal decision, /requirement, idea-stress-test

The skills nearest BoB’s concerns

• handoff: capture state for a fresh agent: describe state rather than instructions, link artifacts instead of copying them, keep the why, present claims as things to verify, never include secrets. BoB’s warp-drive checkpoint and fresh-segment resume do this with persisted state and need no prose hand-off; the session report and the human hand-off record (#3058) cover the rest.
• goal-loop: how to write a /goal prompt: objective, constraints, validation command, stop condition. The same four parts are what loop-primitive declares as data and runs.
• effective-agent-skills: a 13 KB authoring guide (progressive disclosure, lean SKILL.md, push determinism into scripts, validation loops). It agrees with skill-creator and adds nothing BoB’s checks don’t already enforce.
• total-review: two reviewers on different models in parallel, findings merged and triaged, fixes only after the user approves the shortlist. The independent-review idea is already filed from the pm-skills evaluation as #3561; this adds the second model, which BoB, being Claude Code only, cannot use yet.
• risky-changes: before shipping a change to billing, defaults, public API shape or provider parameters, list assumptions, research them, and measure 10–20 realistic cases against the real system. Sound advice; in BoB it is the job of idea-stress-test before a build and of the promotion ladder’s test environment after it. Not taken as a skill.

The pattern taken: a tested command guard (#3563)

global-agent-guardrails describes one denylist of POSIX-ERE patterns, read by a shared hook (deny-dangerous.sh) that every agent calls before a shell command: Claude Code and Codex through PreToolUse on Bash (exit 2 blocks), Cursor through beforeShellExecution, others through small adapters. A test harness (test-guard.sh) runs block and allow cases through each payload shape and must end failed: 0 after any pattern change. Its rules are the useful part:

• Block only the catastrophic: disk wipes, deleting a repo or home, irreversible loss. Recoverable commands (git clean -fdx, rm -rf node_modules) stay allowed.
• It guards against accidents, not a hostile agent; regex can be evaded.
• Fail open when jq or the pattern file is missing, so a broken guard never blocks every command.
• Tell the agent not to retry or work around the block.

BoB’s equivalent today is each automation profile’s permissions.deny list: prefix globs such as Bash(rm -rf *). The supervised profile has none, the lists differ per profile, and nothing tests what they stop. BoB already runs Bash PreToolUse guards through hooks/dispatch.sh, so #3563 adds one declared pattern file and one tested guard to that path. The published repo leaves its own pattern file out, so BoB’s patterns will be its own.

What conflicts with BoB’s conventions

• Decision and idea files. adr-verbatim writes numbered ADR files, and save-idea appends to Markdown backlogs (STARTUP-IDEAS.md and so on) in a separate ideas repository. BoB records decisions and ideas as GitHub issues and forbids local PM files.
• Writing the deployed config. global-agent-guardrails merges hook entries straight into ~/.claude/settings.json, and distribute-skill-to-all-agents authors skills in ~/.agents/skills/ with ~/.claude/skills symlinked to it. In BoB, hook wiring is declared in source and reconciled by deploy.sh, and skills reach BOB_HOME only through a deploy.
• Other harnesses and models. Many orchestration skills drive Codex, Cursor, Grok or GPT, or route through OpenRouter. BoB is Claude Code only until the multi-harness work (#498) lands.
• Paid third-party dependency. Most research skills need a DeepAPI key.

None of these matters for the one pattern taken.

Not verified

• The X post.
• The pattern file the guard reads, which the repository does not publish.
• Any skill’s behaviour: nothing was run.