The problem
Generic agents are fluent about everything and opinionated about nothing.
Ask one for an ADR and it returns a plausible document with five hedged options and no decision. The gap is not model capability — it is that nothing told it what a good ADR refuses to do. Every skilldrop skill carries that judgment with it.
Measured
Every skill is run against the agent without it.
Each acceptance eval runs twice as a real agent session, once with the skill and once without. In the last full run, on claude-sonnet-5, with a skill the agent meets 73% of that skill's checks; without, 53%. A blind judge that sees only the request and the two results preferred the skill's in 62% of pairs, and a second judge in 61%. On the 308 evals written by someone who never read the skill, the two judges preferred it in 58% and 56% of pairs. Some skills were later revised, or admitted, on the verdicts from those same evals, so even that row is partly fitted; what the benchmark showed separates the evals nothing was decided on. The standard tier has since moved to claude-sonnet-5-5; its results there are in what the benchmark showed. Skills that lose are retired: 8 so far, each listed with what to use instead. See every skill's numbers and how they were measured →
Built to be trusted
What a team checks before it adopts a tool, with the evidence one click away.
Stable since 1.0
Commands, flags, skill folders and names, scripts' exit codes and the Action's inputs follow semantic versioning: a break waits for 2.0, after a deprecation.
What stays stable →Scored by OpenSSF Scorecard
Every Action pinned to a commit, CodeQL, workflow and secret scanning on every pull request, and the score published each week.
The scorecard →Releases you can verify
Each release carries the exact npm tarball, a Sigstore-signed record of the build that made it, and an SBOM.
Verify a release →1,312 unit tests
Every skill script is tested on Linux, macOS and Windows, and on Python 3.9, on every pull request. A new script can't land without tests.
How they're run →27 file parsers fuzzed
The scripts that read your exports, reports and configs are fuzzed with Atheris on every change to them, and for longer each week.
What it found →Checked with its own skills
This site's accessibility and performance were reviewed with the catalogue's own skills, and every finding fixed.
The reviews →
See it work
One real run: a launch that is not ready yet.
The input is what a team would paste. The output is what launch-readiness returns, unedited, from its worked example. The rollback is fine, so it does not block, but two failure modes nothing would detect send it back.
Go/no-go for release
orders-v4.2. It splitsorders.addressintoshipping_address_idpointing at a newaddressestable. We followed the migration plan (docs/migrations/orders-address.md): expand and backfill are done in staging, and dual-write is behind the flagorders.address_split. Rollback is "turn the flag off". It was rehearsed in staging on 2026-09-28 (log: ci/run/88213). The p99 target for order reads is 120ms, and the canary at 5% shows 96ms (grafana/d/orders-canary). The existingorders-5xxalert covers this. Release notes are drafted in PR #977. Plan is 5% → 50% → 100% over three days.
launch-readiness returnedLaunch readiness — orders-v4.2 (address split)
Change: release orders-v4.2 · Reach: 5% → 50% → 100% over 3 days · Release owner: Orders tech lead Assessed: 2026-10-01 · Inputs: migration plan, staging rollback log, canary dashboard, PR #977
New failure modes
| # | Failure mode | Who is hurt | Reversible? |
|---|---|---|---|
| F1 | Dual-write diverges: an order is saved with an addresses row that doesn't match orders.address | Customers shipped to the wrong address | Partly. Data can be fixed, parcels already sent cannot |
| F2 | The read path picks the new column before the backfill is complete for that tenant | Customers who see a blank address at checkout | Yes, by turning the flag off |
| F3 | The extra join pushes order reads past p99 under full load | All checkout users | Yes, by turning the flag off |
Checks
| # | Check | Covers | Status | Evidence | Owner |
|---|---|---|---|---|---|
| 1 | Rollback | all | met | Flag-off rehearsed in staging, ci/run/88213. While the old column is still written, flag-off loses no data | Orders on-call |
| 2 | Detection | F1–F3 | not met | orders-5xx catches F3 errors only. Nothing detects F1 divergence or F2 blank reads, because both return 200 | Orders on-call |
| 3 | Response | F1–F3 | not met | There is no runbook entry for divergence. orders-5xx has a runbook entry, but it does not mention the flag | Orders on-call |
| 4 | Quality targets | F3 | met | Canary p99 96ms against a 120ms target at 5%, grafana/d/orders-canary | Orders tech lead |
| 5 | Data & privacy | — | n/a — no new personal data. Address data moves between tables, and retention and processors are unchanged | Migration plan §2 | Data protection lead |
| 6 | Comms | — | met | Release notes drafted in PR #977. No customer-visible change | Orders tech lead |
| 7 | Exposure | — | met | Staged 5% → 50% → 100% behind orders.address_split | Release owner |
Open items
- Row 2: add a reconciliation metric for dual-write divergence, and an alert on any non-zero count, plus an alert on blank-address reads. Orders on-call, before the 5% → 50% step. Hand off to
observability-plan. - Row 3: add a runbook entry for "divergence alert fires" (flag off, then run the reconciler), and add the flag to the
orders-5xxentry. Orders on-call, same deadline. Hand off torunbook-generator.
Verdict
REVISE. Row 2 is not met: the two failure modes that do lasting damage (F1, F2) return 200, so nothing would detect them. Fix rows 2 and 3, then run readiness again.
Read the whole report on the skill page →
Or open one: pages a skill made, live
Each one is a single file that works offline. Click a part to see the source it came from.
Use cases
Start with the job. Meet the packs second.
-
Decide what to build
Product managers · founders · strategists
Turn a direction into a committed, measurable plan: press releases, strategy frameworks, OKRs, business cases, requirements, PRDs, and the metrics that will say whether it worked.
-
Design the system
Architects · tech leads
Commit the shape before the code: decision records, design docs, diagrams, API and data contracts, migration and threat models, and the design of any agentic parts.
-
Build and review software
Software teams · coding agents
The loop around the code itself: supervised implementation, adversarial review panels, test plans, triage, pull request descriptions, a ranked technical-debt register, accessibility and static-analysis passes, and the policy files agents read.
-
Build on the Claude API
Teams building on the Anthropic API
Skills for teams building directly on the Anthropic API — prompt caching strategy, token budgeting, eval generation, and tool-use schema writing.
-
Run and recover the service
SRE · on-call · platform teams
Operate what shipped: runbooks, observability design, release notes, incident communication, and the postmortem that closes the loop.
-
Explain it to decision makers
Anyone presenting to leadership
Get a non-technical audience to a decision: audience profiling, deck outlines and real decks, exec summaries, decision logs, and how-to guides.
-
Govern AI use
AI enablement · engineering leaders · risk owners
Adopt AI deliberately and prove what it is doing: readiness assessment, use-case triage, rollout plans, acceptable-use policy, adoption staging, and usage reporting.
-
Create on-brand collateral
Marketers · designers · anyone making a deck
Make collateral that looks like it came from your team: capture the brand once, then build decks with your logo, colours and fonts.
-
Get a draft ready to ship
Anyone shipping a document
The passes either side of a generator: collect a messy input into a structured brief, critique the argument, and strip the machine artifacts before it leaves the building.
-
Design the experience
Designers · content designers · UX writers
Shape what people see and read before it's built: information architecture, interface copy, content specs, component specs and service blueprints.
-
Research a question
Researchers · analysts · anyone with a question to answer
Take a question to a sourced answer: a research plan with a stopping rule, a cited synthesis of the sources, and competing explanations weighed against the evidence.
-
Define and trust the numbers
Data analysts · analytics engineers
Make the numbers a business runs on mean one thing: SQL reviewed for wrong answers, and dashboards specced before they're built, with every number on them defined.
-
Manage risk and compliance
Compliance, privacy and risk leads
Prepare the material compliance asks for: a data protection impact assessment, a SOC 2 evidence map, a risk register with owners and review dates, and answers to a customer's security questionnaire with the evidence behind each.
-
Build agent skills
Anyone writing skills for an AI agent
Write and review Agent Skills for any tool: a portable SKILL.md with evals, and an audit of an existing skill before it ships.
-
Apply for a job
Anyone applying for a job, in any field
Choose which of an employer's roles to go for, then prepare one application that is tailored and true: an honest fit check against the posting, a resume reworked without adding a claim, a cover letter built only from what you supplied, and the application form's answers drafted from your resume.
How it works
The agent does the work. The decisions you can't undo stay yours.
Five loops run skills in order and stop at a gate before anything moves on. The cheaper the mistake, the more a script decides; the harder it is to undo, the more it waits for you.
- 01 Discoverraw signal to a ratified requirement ratify: you decide
- 02 Designa requirement to a recorded decision ratify: a review panel decides
- 03 Builda requirement to merged code verify: a script decides; decide: you decide
- 04 Releasemerged code to live users, with a way back ready: a review panel decides; go: you decide
- 05 Operatea live service through incidents and what you learn learn: you decide
Plus ship-a-draft, which wraps any document in intake before and critique after. The operating model on one page · Every loop →
Yours after it lands
Files you can read, diff and edit, for you or your whole team.
- A folder you can read
Installing copies plain
How it works →SKILL.mdfolders into the directory your agent already reads. No runtime, no service, nothing to keep running. - Edits that survive updates
Change a skill to fit your codebase.
How it works →skilldrop updatereplaces only files you didn't touch and leaves the new version beside yours as.upstream. - Only what you need
One skill, one pack, one loop, or a whole profile, into Claude Code, Cursor, Kiro, Codex, Copilot or Antigravity.
How it works →
- One setup for everyone
A profile installs a team's packs, loops and reviewer agents in one command.
skilldrop install --profile starterGuide → - Every machine, no per-session step
Bootstrap writes the marketplace into Claude Code settings, so each session already sees the catalogue.
skilldrop bootstrapGuide → - Your own catalogue
Publish your organisation's skills from any repo in the same shape, and install them through the same CLI.
skilldrop install --pack <name> --from <your-repo>Guide →
Install
Start in one command.
16 skills a solution architect reaches for, core included, in one command.
--with-related also pulls the companions it hands off to.
No CLI required. The folder is the whole install.
Skills improve; cp -R never tells you. Files you edited are kept, with the new version beside them as .upstream.