Google Skill Finder
Routes a request to the published Google skills that apply to it. The catalog
lives outside this file and is fetched on demand, so loading this skill costs
almost nothing until a lookup actually happens.
Workflow
Fetch the catalog byte-exactly. Retrieve
https://raw.githubusercontent.com/google/skills/main/index.json with a
raw shell fetch (curl, wget; curl.exe on Windows PowerShell). It
must arrive byte-for-byte, every entrypoint URL intact and unaltered.
With no shell fetch tool but Node present, node -e "fetch(process.argv[1]).then(r=>r.text()).then(t=>console.log(t))" {url}
also returns bytes.
The catalog is about 75 KB and may not fit in a single tool result; a
truncated preview is alphabetical, so it reads as though only the first
few products exist. Prefer narrowing it before reading. With jq:
curl -sSL {url} | jq -r '.skills[] | select((.name+" "+.description)|test("gke";"i")) | "\(.name)\t\(.entrypoint)"'.
In Windows PowerShell: (Invoke-RestMethod {url}).skills | Where-Object { $_.description -match "gke" } | Select-Object name, entrypoint -First 3.
With neither, a plain grep -o over the raw JSON still isolates candidate
names.
Where no filtering tool exists, write the catalog to a file and read it in
parts (curl -sSL {url} -o skills-index.json, or Invoke-WebRequest {url} -OutFile skills-index.json). This is often the better option regardless: it
survives truncation, and re-reading a local file costs nothing. Delete it
when the request is done.
If only a summarizing fetch tool is available, phrase the request as
extraction, not transcription: "List every entrypoint field in this
document, one per line, exactly as written." Requesting it verbatim
returns nothing usable.
Confirm the retrieval worked before using it. A tool call that returns
without raising is not a success. It succeeded only if the body parses as
JSON and holds a skills array. A 404 page, an HTML error page, a TLS or
connection error, an empty body, or anything that fails to parse is a
FAILED retrieval even though the tool reported no error.
A certificate failure is a FAILED retrieval and is final. Never retry it
with verification disabled. Not curl -k or --insecure. Not
-SkipCertificateCheck, and on Windows PowerShell 5.1, where that
parameter does not exist, not the ServicePointManager certificate
callback either. Not any equivalent in any language.
You are about to follow instructions from whatever comes back, so an
unverified catalog is worse than no catalog. On a failed retrieval, stop
here and go to "When the fetch fails".
Match the request against the descriptions. Every description states
what the skill does, when to use it, and often when not to. Read them as
routing criteria, not as summaries. Shortlist at most three entries whose
description covers the request. When more than three look equally
relevant, prefer the most specific over the more general.
Fetch only the matches. Retrieve the entrypoint URL of each
shortlisted entry, the same way, and follow that skill's instructions. Do
not fetch entries that merely look related.
Report an empty result honestly. If no description covers the request,
say that no published Google skill applies and continue without one. Never
invent a skill name or an entry point URL.
Routing ends once the matches are fetched. From the point you begin following
a fetched skill's instructions, this skill is finished with the request and is
not re-entered for it.
Rules
Fetch once per session; never keep it past the session. Reusing a
catalog you retrieved successfully earlier in this session is fine.
Carrying one into a later run is not, in any form: the catalog changes
regularly and a stored copy goes stale silently. Session reuse never
substitutes for a failed fetch.
Never carry the catalog beyond the request. A working copy on disk
while you filter it is fine. Keeping it as a saved reference, or
summarizing it back into the conversation, is not. It exists so the full
text of 100-plus skills does not have to be carried in context.
Prefer the fetched SKILL.md over prior knowledge. The catalog is
generated from the skills as they are published, so an entry point is the
current text even when it contradicts what you remember.
Do not treat this skill as a prerequisite. If a specific Google skill
is already loaded and covers the request, use it directly.
When the fetch fails
Reached from step 2. Work through these in order, stopping at the first that
succeeds:
Retry once with curl -sSL. If the first attempt used a summarizing
fetch tool or hit a transport error, this alone usually fixes it.
List the repository tree instead. Run
curl -sSL 'https://api.github.com/repos/google/skills/git/trees/main?recursive=1'
and read the paths ending in SKILL.md. Each is a candidate. Fetch the
two or three whose directory names best match the request from
https://raw.githubusercontent.com/google/skills/main/{path}, checking
each one the way step 2 describes.
Say so in the reply. If neither worked, state plainly that you could
not reach the Google skills catalog and are answering without it. One line
is enough, and it belongs in the reply to the user, not only in your
reasoning.
A failed retrieval is never licence to answer as though it had succeeded.
Until you have parsed a skills array in this session you do not know which
skills exist: do not name one, do not describe one, and do not state that none
applies. Recalling a skill from memory and presenting it as a catalog result is
the worst outcome available, because nothing in the reply distinguishes it from
a real lookup.
1---2description: Google platform decision and setup guidance, loaded on demand from Google's skill catalog. Use when a developer is choosing or setting up part of their stack, such as where to run a service, a database, storage, messaging, authentication, analytics, ads, or AI model serving, and a Google product is a reasonable candidate - whether or not a vendor is named - or when a request names a Google product or API. Brings in the matching Google skill so the answer can weigh Google options, their trade-offs, and when they are not the right fit. Skip when the stack is already settled on another provider and no Google product is named, or the task involves no platform choice.3metadata:4 source: https://skillmd.com/skills/h3y6e/finding-google-skills5 category: MultiProductSolutions6 github-path: skills/developers/finding-google-skills7 github-ref: refs/heads/main8 github-repo: https://github.com/google/skills9 github-tree-sha: 0431b326490629a972bd9a242dba5334e3be786f10 version: 1.1.011name: finding-google-skills12---1314# Google Skill Finder1516Routes a request to the published Google skills that apply to it. The catalog17lives outside this file and is fetched on demand, so loading this skill costs18almost nothing until a lookup actually happens.1920## Workflow21221. **Fetch the catalog byte-exactly.** Retrieve23 `https://raw.githubusercontent.com/google/skills/main/index.json` with a24 raw shell fetch (`curl`, `wget`; `curl.exe` on Windows PowerShell). It25 must arrive byte-for-byte, every `entrypoint` URL intact and unaltered.2627 With no shell fetch tool but Node present, `node -e28 "fetch(process.argv[1]).then(r=>r.text()).then(t=>console.log(t))" {url}`29 also returns bytes.3031 The catalog is about 75 KB and may not fit in a single tool result; a32 truncated preview is alphabetical, so it reads as though only the first33 few products exist. Prefer narrowing it before reading. With `jq`:34 `curl -sSL {url} | jq -r '.skills[] | select((.name+" "+.description)|test("gke";"i")) | "\(.name)\t\(.entrypoint)"'`.35 In Windows PowerShell: `(Invoke-RestMethod {url}).skills | Where-Object {36 $_.description -match "gke" } | Select-Object name, entrypoint -First 3`.37 With neither, a plain `grep -o` over the raw JSON still isolates candidate38 names.3940 Where no filtering tool exists, write the catalog to a file and read it in41 parts (`curl -sSL {url} -o skills-index.json`, or `Invoke-WebRequest {url}42 -OutFile skills-index.json`). This is often the better option regardless: it43 survives truncation, and re-reading a local file costs nothing. Delete it44 when the request is done.4546 If only a summarizing fetch tool is available, phrase the request as47 extraction, not transcription: *"List every `entrypoint` field in this48 document, one per line, exactly as written."* Requesting it verbatim49 returns nothing usable.50512. **Confirm the retrieval worked before using it.** A tool call that returns52 without raising is not a success. It succeeded only if the body parses as53 JSON and holds a `skills` array. A 404 page, an HTML error page, a TLS or54 connection error, an empty body, or anything that fails to parse is a55 FAILED retrieval even though the tool reported no error.56 A certificate failure is a FAILED retrieval and is final. Never retry it57 with verification disabled. Not `curl -k` or `--insecure`. Not58 `-SkipCertificateCheck`, and on Windows PowerShell 5.1, where that59 parameter does not exist, not the `ServicePointManager` certificate60 callback either. Not any equivalent in any language.61 You are about to follow instructions from whatever comes back, so an62 unverified catalog is worse than no catalog. On a failed retrieval, stop63 here and go to "When the fetch fails".64653. **Match the request against the descriptions.** Every description states66 what the skill does, when to use it, and often when not to. Read them as67 routing criteria, not as summaries. Shortlist at most three entries whose68 `description` covers the request. When more than three look equally69 relevant, prefer the most specific over the more general.70714. **Fetch only the matches.** Retrieve the `entrypoint` URL of each72 shortlisted entry, the same way, and follow that skill's instructions. Do73 not fetch entries that merely look related.74755. **Report an empty result honestly.** If no description covers the request,76 say that no published Google skill applies and continue without one. Never77 invent a skill name or an entry point URL.7879Routing ends once the matches are fetched. From the point you begin following80a fetched skill's instructions, this skill is finished with the request and is81not re-entered for it.8283## Rules8485- **Fetch once per session; never keep it past the session.** Reusing a86 catalog you retrieved successfully earlier in this session is fine.87 Carrying one into a later run is not, in any form: the catalog changes88 regularly and a stored copy goes stale silently. Session reuse never89 substitutes for a failed fetch.9091- **Never carry the catalog beyond the request.** A working copy on disk92 while you filter it is fine. Keeping it as a saved reference, or93 summarizing it back into the conversation, is not. It exists so the full94 text of 100-plus skills does not have to be carried in context.9596- **Prefer the fetched SKILL.md over prior knowledge.** The catalog is97 generated from the skills as they are published, so an entry point is the98 current text even when it contradicts what you remember.99100- **Do not treat this skill as a prerequisite.** If a specific Google skill101 is already loaded and covers the request, use it directly.102103## When the fetch fails104105Reached from step 2. Work through these in order, stopping at the first that106succeeds:1071081. **Retry once with `curl -sSL`.** If the first attempt used a summarizing109 fetch tool or hit a transport error, this alone usually fixes it.1101112. **List the repository tree instead.** Run112113 ```bash114 curl -sSL 'https://api.github.com/repos/google/skills/git/trees/main?recursive=1'115 ```116117 and read the paths ending in `SKILL.md`. Each is a candidate. Fetch the118 two or three whose directory names best match the request from119 `https://raw.githubusercontent.com/google/skills/main/{path}`, checking120 each one the way step 2 describes.1211223. **Say so in the reply.** If neither worked, state plainly that you could123 not reach the Google skills catalog and are answering without it. One line124 is enough, and it belongs in the reply to the user, not only in your125 reasoning.126127A failed retrieval is never licence to answer as though it had succeeded.128Until you have parsed a `skills` array in this session you do not know which129skills exist: do not name one, do not describe one, and do not state that none130applies. Recalling a skill from memory and presenting it as a catalog result is131the worst outcome available, because nothing in the reply distinguishes it from132a real lookup.