How to Make Your Agent Better at Workflow
A skills.sh shortlist for brainstorming, planning, TDD, review, and debugging, with a download link and install command for each skill.
Critique
GUIDES / AGENT SKILLS
Better workflow
Brainstorm, plan, TDD, review.
critique.sh
Agent harness · BYOA · Composer 2.5
Install the Superpowers sequence from skills.sh: `using-superpowers`, `brainstorming`, `writing-plans`, `test-driven-development`, `requesting-code-review`, and `systematic-debugging`. Pair that with `find-skills` so the agent can fetch more skills instead of improvising. Skills shape the author loop. They do not replace an independent finish pass.
Most agents start coding in the first tool call. The expensive failure is not a syntax error. It is forty minutes of implementation against the wrong spec, followed by a test that documents the bug.
Workflow skills are gates. Brainstorming is a gate against coding. Planning is a gate against holding the entire feature in context. TDD is a gate against untested code. Review is a gate against the author grading itself. If those gates are optional, the agent will skip them whenever the prompt sounds urgent.
What skills.sh actually installs
skills.sh is the public directory for Agent Skills: folders of instructions, and sometimes scripts, that a coding agent can load when the task matches. The package manager is npx skills. The discovery skill is `find-skills` from vercel-labs/skills (3.4M installs).
A skill is not a model upgrade. It is a procedure the agent is supposed to follow: which files to read, which checks to run, which output shape to return, and when to stop. That is why a mediocre prompt plus a good skill often beats a long chat with no procedure.
The loop you actually want
`obra/superpowers` is the coherent pack for that loop. Matt Pocock’s listing is the other high-install workflow cluster: `implement` (556.7K), `code-review` (553.3K), and `diagnosing-bugs` (604.4K). Pick one pack as the default so skills do not contradict each other.
- 1Has anyone agreed what to build?brainstorming. No files, no scaffolding, no “quick spike” that becomes the architecture.
- 2Is the design approved but the work is still a blob?writing-plans. Demand file paths, verification commands, and tasks small enough to fail independently.
- 3Are you implementing behavior?test-driven-development. Watch the test fail. If it passes immediately, you tested the wrong thing.
- 4Did a unit of work just land?requesting-code-review. Give the reviewer SHAs and the spec, not the author transcript.
Six workflow skills worth installing
using-superpowers is the on-ramp. Without it, the rest of the pack is a folder the agent never opens. The other five skills are the actual gates. subagent-driven-development (204.9K) is the optional accelerator once the plan exists: one fresh subagent per task, then spec review, then quality review.
| Skill | Use it for | Installs | Download |
|---|---|---|---|
using-superpowers | Load the Superpowers workflow so brainstorming, planning, TDD, and review fire as a sequence, not as optional extras. | 243.6K | Download on skills.sh |
brainstorming | Refuse to write code until the design is written down and the human has signed off. | 363.1K | Download on skills.sh |
writing-plans | Break an approved design into small TDD tasks with file paths, commands, and expected output. | 248.7K | Download on skills.sh |
test-driven-development | Write a failing test, watch it fail, then write the minimum code that passes. Delete code written before tests. | 226.4K | Download on skills.sh |
requesting-code-review | Dispatch a reviewer subagent with SHAs and requirements, never the author session’s history. | 228.3K | Download on skills.sh |
systematic-debugging | Investigate failures with evidence instead of restarting the feature from a fresh prompt. | 259.2K | Download on skills.sh |
Open the download page
Each card goes to the skills.sh listing. From there you can copy the install command or inspect the SKILL.md.
obra/superpowers.obra/superpowers.obra/superpowers.obra/superpowers.obra/superpowers.obra/superpowers.discover
npx skills add https://github.com/vercel-labs/skills --skill find-skills
npx skills find tdd review workflowBrowse the leaderboard, then search by job rather than by brand.
using-superpowers
npx skills add https://github.com/obra/superpowers --skill using-superpowershttps://skills.sh/obra/superpowers/using-superpowers
brainstorming
npx skills add https://github.com/obra/superpowers --skill brainstorminghttps://skills.sh/obra/superpowers/brainstorming
writing-plans
npx skills add https://github.com/obra/superpowers --skill writing-planshttps://skills.sh/obra/superpowers/writing-plans
test-driven-development
npx skills add https://github.com/obra/superpowers --skill test-driven-developmenthttps://skills.sh/obra/superpowers/test-driven-development
requesting-code-review
npx skills add https://github.com/obra/superpowers --skill requesting-code-reviewhttps://skills.sh/obra/superpowers/requesting-code-review
systematic-debugging
npx skills add https://github.com/obra/superpowers --skill systematic-debugginghttps://skills.sh/obra/superpowers/systematic-debugging
Write your own skill only after you have a repeatable failure
`skill-creator` from Anthropic (381.5K installs) is the right tool when your team keeps re-explaining the same procedure. Do not start there. Start by installing a public skill that already encodes the loop. Create a private skill when you can name the failure: “agents skip the billing invariant,” not “we should have more skills.”
critique does the same for an author you already use. See the critique-review skill writeup for the portable review skill across Claude Code, Codex, Hermes, and OpenCode.Skills improve the author. They do not finish the change.
A workflow skill can still grade its own homework. Dispatching a reviewer subagent is better than no review. It is not the same as a controller-owned finish pass.
If Claude Code, Codex, Cursor, or OpenCode is already the author, keep it. Install the skills above, then run an independent finish pass with the critique sidecar when a unit of work is ready. If you want Critique to be the author, use CritiqueCode: it implements, then the controller forces review and verified repair before promotion. Those are different binaries. Do not mix the jobs.