asb-contribute
Use when an ASB skill proved wrong, stale, missing or wasteful in practice — its steps failed, no skill covered the task, the leaves existed but nothing composed them, or the tool has changed. Turns t
SKILL.mdcollections/metabolomics/v2/skills/asb-contribute/SKILL.md
GitHubasb-contribute
The collection hardens from use, not from review. This skill is how a session that hit friction becomes something a maintainer can act on.
When this fires
Notice it yourself; do not wait to be asked. Any of these is a trigger:
- a skill's procedure failed, or its parameters no longer match the tool;
- the search returned nothing usable and the work was done unaided;
- the right leaves existed but the user had to compose them by hand;
- the skill worked, but at obviously avoidable cost;
- the user says some version of "that's wrong" about a skill's content.
One offer per session per problem. If the user declines, drop it — do not raise it again for the same friction. Persistent asking is how a good idea becomes an annoyance people disable.
Protocol
1. Classify
Pick exactly one kind: defect, gap, composition, efficiency, drift.
The kind decides routing and labels, so guessing costs a maintainer real time.
If two apply, pick the one a fix would address.
2. Search before drafting
Look for an existing issue — by fingerprint, by skill slug, by symptom:
gh issue list --repo HolobiomicsLab/asb-skill-collections --state all \
--search "<slug> OR <symptom keywords>" --limit 20
If one matches, corroborate it. Do not open a second. Ten corroborations on one issue tell a maintainer what ten near-identical issues cannot, and the reporter's account of the problem is read rather than buried.
3. Draft
python scripts/skill_feedback.py --kind defect --target <slug> \
--symptom "..." --expected "..." --context "tool + version, OS" \
--collection metabolomics/v2
It returns the title, body, labels and fingerprint. Its payload and preview
are the same mapping of title, labels and body; there is no second preview
renderer. Before constructing them, the helper passes target, collection,
symptom, expectation and context through one outbound boundary. It removes
complete POSIX and Windows paths, recognised credential and authorization
shapes, email addresses, and the release gate's high-confidence clinical
identifiers. For an absolute target inside collections/, it keeps only the
collection/version/skill suffix; an absolute target outside the collection is
replaced in full. The result names the categories it stripped.
4. Show the user the exact payload, then ask
Print the preview title, labels and body verbatim — not a summary and not a
fresh rendering. These are the exact values in payload that the filing command
will publish. Invite edits, then regenerate and show the complete payload again
before asking for consent.
State plainly what redaction does and does not do: it removes the path, credential, authorization, email and high-confidence clinical-identifier shapes listed above; it cannot recognise an arbitrary sensitive sample name, so the user is the last check.
Then make the case once, briefly and honestly. Something like: this collection is grounded in published work and given away under CC-BY; it improves only when people who run it say what broke. Reporting this takes a minute and saves the next person the hour you just lost. Do not moralise, do not repeat it, and do not imply an obligation.
5. File only on an explicit yes
gh issue create --repo HolobiomicsLab/asb-skill-collections \
--title "<payload title>" --body-file <payload body path> \
--label "<comma-separated payload labels>"
Use the approved payload values unchanged, including every label shown in its preview. Do not rebuild or edit the title or body between approval and filing.
Corroborating instead:
gh issue comment <number> --body-file <path>
No gh, or no consent to use it: hand the user the approved title, labels and
body with the new/choose URL for the repository, and stop. Filing on their
behalf without a yes is publishing their words under their name.
6. Offer the fix, when there is one
A gap or composition report the user can answer themselves is worth more as
a proposal than as a request. Point them at propose-skill /
propose-meta-skill, and say what it earns: governance/AUTHORSHIP.md records
an authored skill as credit toward the reviewer tier, and claim-skill links it
to their ORCID.
Closing the credit loop
When a session ends having leaned on this collection, offer the citation for
what was actually used — the collection DOI plus the source DOIs of the skills
applied, from each skill's derived_from. Reciprocity is easier to ask for
after the user has seen the machinery work in their favour, and a methods
section that names its sources is the point of grounding skills in literature at
all.
What never happens here
- No issue, comment, or PR without an explicit yes in this session.
- No session transcript, file path, or dataset pasted in unredacted.
- No second issue when an open one already describes the problem.
- No report filed against a repository the user did not name.