vet-and-adopt
Skill0 starsChanged 7 days agoMITView on GitHub
Use for each item in an adoption plan, and whenever someone wants to take a skill, hook, agent or practice from this kit or from any outside source into their own Claude setup. Scans shipped files with SkillSpector first, triages the findings, recommends SKIP / FOLD / ADAPT / ADOPT against what they already have, applies only what they decide, and records the decision. Trigger on "adopt <item>", "next item in the adoption plan", "vet this skill", "is this skill safe", "take this into my setup".
In this repository5 other files it gives its agents.
- AGENTS.mdAGENTS.md
- templates/AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- drill-to-adoptSkill
- setup-workspaceSkill
The file
We split this skill into the rules it gives an agent
8 rules. About 0 are common, 3 are distinctive, and 5 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 49
Run its Check. Tick the row only after the check has passed, and fill in the row's Check cell.
- Very specific: line 62
Read the tailoring for it in their adoption/spec.md, if there is one.
- Distinctive: line 78
Before the first scan result of a sitting, say once how loud the scanner is: "SkillSpector reads every file and flags anything that could …
- Distinctive: line 86
Anything else: stop. Show the score and each finding with your reading of it: a real risk, or a false positive and why (quote the line).
- Very specific: line 90
Read the files yourself as well.
- Very specific: line 110
Fit, not copy: Write every adopted file in the adopter's words, for their tools and scale.
- Very specific: line 111
Never copy a studio sentence, term or path.
- Very specific: line 128
MCP servers: add one with claude mcp add at the narrowest scope that works, local first, and pass secrets as environment variables or …
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
vet-and-adopt
One item at a time: scan, decide, apply, record, check. The scanner is a gate, not a judge. A clean
scan does not mean the item fits, and a noisy scan is often false positives; both need your reading.
The decision is always the adopter's. drill-to-adopt's "Voice and honesty" and "Rules for every
question" hold here in full: one question per message, approvals included; the recommended option
first, then Not sure, and Skip for now where it differs; what a command does, in plain words, before
the command.
Taking the next row of the plan
With an adoption plan (adoption/plan.md in their setup), work from it, wave 1 only in a first
sitting:
- Take the first row without a tick in its Done column.
- When row 1, SkillSpector with the scan-first rule, is not ticked, do it first, from its card,
cards/skillspector.md: the install lines for their system as!commands (uv, then SkillSpector) when it is missing; when Detect found it, you run the two checks yourself, since they only print:skillspector --versionprints a version, and a scan of one kit skill prints a verdict. Nothing else is scanned, so nothing else is adopted, before those two checks pass. The rule line may wait for the row that creates the front door; row 1 counts as done for this gate once its two checks pass. - "Scan what you already have", when the plan has it: say in one line that you are scanning
their installed items, read-only, then scan each one it lists with SkillSpector from where it is
installed, and triage as in section 2. Change nothing. Tell them the result in one line per item.
Record each result in
adoption/decisions.md. A real finding becomes a question to them, never an automatic removal. - For every item, open its card,
cards/<id>.mdin the kit. A card question their spec has no answer to is asked now, in full, before anything else; a menu tick is not an answer. Then follow the Steps in order, each you: step handed over (drill-to-adopt, "Steps only they can do") and each Claude: step shown before it runs. - Record the decision as soon as the item is applied or staged (section 5), with Check "pending". A SKIP gets its row too, with Check "n/a".
- Run its Check. Tick the row only after the check has passed, and fill in the row's Check cell. A failed check leaves the row unticked; say what failed and what to look at next.
Without a plan (an item someone names directly), the steps below still apply in the same order.
0. Identify the item
- From the kit:
catalog/skills/<id>/,catalog/agents/<id>.md, a file intemplates/, or a practice frommodules/<nn>/README.md. Read its card (cards/<id>.md), itsSOURCE.mdand its module page. A domain iscatalog/packs/<id>/, with itsSOURCE.mdin that folder and its card atcards/pack-<id>.md. - From outside: a folder, a
SKILL.md, a.zipor a git URL; or a plugin, an MCP server or a hook, which run inside a session too and are scanned from their source repository the same way. - Read the tailoring for it in their
adoption/spec.md, if there is one.
1. Scan, when the item ships files
A static scan, no API key, no cost:
skillspector scan <path-or-git-url> --no-llm --format markdown --output <a temp folder>/scan.mdOn Windows put PYTHONUTF8=1 in front, which turns on Python's UTF-8 mode; without it the report
crashes on the console's encoding (docs/facts.md, "Installing SkillSpector"). Keep the paid LLM
pass off unless the adopter chooses it. A practice with no files skips the scan.
2. Triage
Before the first scan result of a sitting, say once how loud the scanner is: "SkillSpector reads every file and flags anything that could be misused, so most of its findings are false alarms; the score says how much it flagged, not how dangerous the item is."
- SAFE: continue.
- A kit item that is not SAFE: compare its findings with the kit's own triage
(
tools/export/scan-triage.mdand the item'sSOURCE.md) and show "known findings: N, already read in the kit's triage; new: M". Ask them to weigh only a new finding, as below. - Anything else: stop. Show the score and each finding with your reading of it: a real risk, or a false positive and why (quote the line). Continue only when the adopter has read them and says so; never ask them to wave through a red score unexplained. Prompt injection, data exfiltration, or a network call they did not expect is a hard stop.
- Read the files yourself as well. A scanner can miss a real egress. When the studio reviewed an
outside framework, the scan scored it 100/100 critical on false positives and missed its
telemetry, which was on by default (
examples/outside-framework-review.md).
3. Compare with what they have
List what in their setup already does this job (from round 1 of the drill, or look). Recommend one decision, with one line of why:
- SKIP: it does not fit their work, or they already have something better.
- FOLD: take the idea into something they already have: an existing skill, their instruction file, a Project's instructions.
- ADAPT: take it, rewritten for their setup (their paths, tools and rules), with the
studio-specific steps removed. Those steps are marked in the module page and in
SOURCE.md. - ADOPT: take it as it is.
The adopter decides. Apply nothing before they do.
4. Apply
- Fit, not copy. Write every adopted file in the adopter's words, for their tools and scale.
Never copy a studio sentence, term or path. A module's
(studio-specific: ...)notes are never carried over. A practice defaults to ADAPT. Take the size they chose on the item's card, and nothing beyond it. - Claude Code: skills go in
<their repo>/.claude/skills/<id>/for one project, or~/.claude/skills/<id>/for every project; agents in.claude/agents/; instruction text in theirCLAUDE.mdorAGENTS.md. Keep the upstreamLICENSEwith any adapted third-party skill. - A domain (a pack in
catalog/packs/<id>/): scan the whole folder first (step 1), then copy by these rules, each file shown before it is saved:- the entry
SKILL.mdand theskills/folder to<scope>/skills/<id>/, kept together, because the entry points to the files inskills/by relative path; - each
stubs/<name>/SKILL.mdto<scope>/skills/<name>/SKILL.md(a stub is a one-line skill that sends Claude into the pack, for a name a person types or a skill with tool limits); - each file in
agents/to<scope>/agents/, only the ones the adopter took on the card. Then check it (the card's Check). The pack's own text is the studio's, and its examples are the studio's too: read them with the adopter and swap any that name a tool or a path they do not use. On claude.ai a pack goes in as a zip built bynode tools/upload-bundle.mjs <id> --out <folder>: the entry and the skills, with no agents.
- the entry
- MCP servers: add one with
claude mcp addat the narrowest scope that works, local first, and pass secrets as environment variables or${VAR}references, never as text in a shared.mcp.json(docs/facts.md, "Claude Code: MCP servers"). A sign-in the server needs is handed over as a!command for them to type. - claude.ai: instruction text goes into a Project's instructions, reference documents into the
Project's knowledge, skills through the upload described in
docs/facts.md("claude.ai: custom skills and the built-in document skills"). - Respect the data rules in their spec. Nothing company-internal goes where their rules forbid it, such as a personal account or an outside remote.
- Show them the change before you save it.
5. Record
Append one row to their adoption/decisions.md (format in templates/adoption/decisions.md) as soon
as the item is applied or staged, before its check: date, kit version (the top section of
CHANGELOG.md), item, decision, where it went, why, the scan result, Check "pending", and the undo
(the card's Undo, written for their setup). When the check has run, replace "pending" with its
result and the line that showed it. One row per item, a folded item included.
6. Outside items
The same steps apply to a skill from anywhere. Name its upstream and its licence in the decisions row, and keep its licence file with any copy.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

