Last updated: November 9, 2026
Short answer: Addy Osmani's agent-skills repo is full of stop rules. interview-me stops asking when it can predict your reaction to its next three questions. spec-driven-development ends its turn after the spec. doubt-driven-development caps adversarial review at three cycles. code-review-and-quality treats zero findings as a valid result. Most of them come with a condition you can check.
We run Skillselion, an independent directory of skills and MCP servers ranked by real installs. It is not affiliated with Anthropic. Adapted from our Addy Osmani's agent skills guide on Skillselion. Every quote below was checked against the current SKILL.md on GitHub, and every install count comes from the live listing page on November 9, 2026.
A lot of coding-agent mistakes are timing mistakes. The agent starts building before it understands the ask, or it keeps going after it should have checked with you, or it reviews its own work and finds nothing because it wrote it. So we read five SKILL.md files from addyosmani/agent-skills (more than 114,000 GitHub stars on November 9, 2026) and pulled out one thing from each: the exact point where the skill tells the agent to stop.
Why does Claude Code start coding before it understands the task?
Often because nothing told it to wait. The router skill, using-agent-skills (67,092 installs), fixes that at the session level. using-agent-skills router skill listing, source on GitHub.
It does two jobs. First, it maps the task to a phase and names the skill to load: interview-me when you don't know what you want yet, spec-driven-development for a new feature, code-review-and-quality before merge. Second, it sets behaviors that apply under every other skill. The one that matters here is confusion handling. When the agent hits conflicting requirements, the instruction is "STOP. Do not proceed with a guess." It then has to name the specific confusion and wait.
When the ask is unclear but not contradictory, it uses a softer gate. The agent lists its assumptions in a numbered block that ends with "Correct me now or I'll proceed with these." That line gives you one chance to correct a wrong guess before it turns into code. It costs you one reply.
How does interview-me know when to stop asking questions?
interview-me (65,507 installs) runs before any spec exists. interview-me skill listing, source on GitHub.
The agent opens with a one-sentence hypothesis and a confidence number, then asks one question at a time with its own guess attached. The stop condition is a single test:
"Can I predict the user's reaction to the next three questions I would ask?"
If yes, the agent writes back a restate with six lines: outcome, user, why now, success, constraint, out of scope. The skill calls that last line non-negotiable, and gives the reason: "Half of misalignment is silent disagreement about what is not being built."
The confirmation gate is strict. It needs an explicit yes. "Sounds good" does not count, and the skill says so: "Silence isn't confirmation." "Whatever you think is best" also fails, because the user is delegating. The agent is told to re-ask with two concrete options. Once you do say yes, the agent delivers the statement of intent, offers next steps, and stops. In the skill's words: "STOP YOUR TURN IMMEDIATELY." No tools, no spec, no code in that turn.
There is also a floor. If several rounds pass and the agent still cannot predict your answers, it should say so and suggest stepping back. And the skill rules itself out of CI, scheduled runs and /loop: "This skill needs a live, responsive user."
If you already use Matt Pocock's grill-me, the two fit together. grill-me describes itself as "A relentless interview to sharpen a plan or design," so it works on a plan you already have. Run interview-me first. Our map of Matt Pocock's skills, grill-me included covers that side.
How does a skill make Claude Code wait for your review between spec and code?
spec-driven-development (73,694 installs) ends its turn after the spec, so a hard stop sits between the spec and any code. spec-driven-development skill listing, source on GitHub.
The workflow has four gated phases (Specify, Plan, Tasks, Implement), and the rule between them is plain: "Do not advance to the next phase until the current one is validated." After the spec is saved, the agent summarizes it, lists open questions, asks you to approve, and ends its turn. "Planning starts only after the human approves the spec in a later turn."
Two details are worth stealing even if you never install the skill. The spec template includes a three-tier Boundaries section: always do, ask first, never do. And vague requirements get rewritten as success criteria. "Make the dashboard faster" becomes targets like LCP under 2.5s and CLS under 0.1, followed by a question asking whether those are the right targets. For other ways to run this phase, see our spec-driven development skills guide.
How do you stop an agent from agreeing with itself?
doubt-driven-development (64,915 installs) has the strictest loop bound in the repo. doubt-driven-development skill listing, source on GitHub.
It applies only to non-trivial decisions. The skill defines those with a checklist: the change touches branching logic, crosses a module or service boundary, claims a property the compiler cannot check (thread safety, idempotence, ordering), depends on context a future reader cannot see, or cannot be undone. Renames and formatting are out. The skill puts it bluntly: "If you doubt every keystroke, you ship nothing."
For a decision that qualifies, the agent writes a short CLAIM, extracts the smallest reviewable ARTIFACT plus its CONTRACT, and hands those to a fresh-context reviewer with an adversarial prompt. What it withholds matters most:
"Pass ARTIFACT + CONTRACT only. Do NOT pass the CLAIM."
If the reviewer sees your conclusion, it tends to agree with it. Without the claim, it has to check the artifact against the contract on its own.
The stop rule is a hard bound. The loop ends when findings turn trivial, when you say ship it, or after three cycles. The agent does not "grind a fourth alone." If three cycles still surface real issues, the agent escalates to you, and if the artifact is too big for three cycles, it goes back and splits it: "Do not lift the bound."
In interactive sessions the skill also makes the agent offer a cross-model second opinion (Gemini CLI, Codex CLI, or a manual paste) on every cycle. You can decline. What the agent cannot do is skip the offer without telling you. "Skipping is fine; silent skipping is not." The skill wants that external CLI run in a read-only sandbox with the prompt piped through stdin, since the artifact itself might contain instructions.
What does code-review-and-quality check that a normal review misses?
code-review-and-quality (83,040 installs) has the most installs of the repo's skills in our catalog. code-review-and-quality skill listing, source on GitHub.
It reviews five axes (correctness, readability, architecture, security, performance) and sizes changes on a simple scale: about 100 changed lines is good, about 300 is acceptable for one logical change, and about 1,000 is "Too large. Split it."
Two stop rules here are easy to overlook. The first protects you from a reviewer that fills space. Before reporting, the agent checks each finding against the callers, guards and tests that reach the changed code, and the skill says: "Zero findings is a valid result; do not invent one." The second is a test you can run yourself. Instead of reading tests to judge coverage, the agent flips one condition the change adds (drop a negation, swap && for ||), runs the suite, and restores the file. "A mutation that stays green is a finding: name the test case that is missing."
There is also a rule for dead code: list it, then ask before deleting. To compare this skill with other reviewers, see our roundup of code review skills for Claude Code.
How do you install Addy Osmani's agent skills in Claude Code?
One skill at a time:
npx skills add addyosmani/agent-skills --skill interview-me
Or the whole set, which the router needs, since it can only route to skills you have:
npx skills add addyosmani/agent-skills
The repo also ships as a Claude Code plugin:
/plugin marketplace add addyosmani/agent-skills
/plugin install agent-skills@addy-agent-skills
One catch from the README: a per-skill npx install copies only skills/<name>/, not the shared references/ folder. The skills still work, but paths to the shared checklists are unavailable. doubt-driven-development points at references/orchestration-patterns.md, and code-review-and-quality points at the security and performance checklists. If you want those, install the whole repo. The README's install-all line says 25 skills, including constraint-driven-development, which the router sends you to when "No quality bar written down?"
The addyosmani/agent-skills marketplace page lists the skills it has indexed so far, with live install counts. If you are choosing between workflow packs, the side-by-side comparison of Superpowers and agent-skills shows how the two differ.
FAQ
Do these skills slow Claude Code down?
On ambiguous or risky work, yes, on purpose. Both interview-me and doubt-driven-development list "speed over verification" as a reason not to run, and both skip mechanical edits like renames.
Can I use interview-me in CI or a scheduled agent?
No. The skill excludes non-interactive contexts. If the ask is underspecified there, it should flag a blocker instead of guessing.
What is the difference between doubt-driven-development and code-review-and-quality?
Timing. Doubt-driven reviews single decisions while the work is in progress. code-review-and-quality reviews the finished change before merge. The doubt skill tells you to use both.
Which one should I install first?
If you only take one, take code-review-and-quality: 83,040 installs on November 9, 2026, and it pays off on every change. If your problem is agents building the wrong thing, start with interview-me.
Install counts: live catalog, refreshed daily from skills.sh, GitHub and MCP registries, ranked by real installs. Figures as of November 9, 2026.
Top comments (0)