| name | source-currency-auditor |
| description | Audit the library's skills that cite external sources against a known-good source list, and flag every citation older than a stated threshold (N months) — or superseded/broken — for re-verification. The broad currency sweep: inventory external-source citations across all skills (standards editions, external docs/URLs, research, tool versions, prices/limits, dated facts), check each against a known-good list, flag stale/broken/superseded citations with the reason, and prioritize by how load-bearing and volatile the claim is. A periodic staleness gate, not an edition-specific tracker. Reports for re-verification; verifies nothing itself and changes nothing. Use when auditing citation freshness across the library, setting a currency threshold, or periodically sweeping for stale external references. Do NOT use to track specific framework EDITIONS and their deltas (framework-edition-tracker), propose mapping edits (framework-mapping-refresher), or review a skill's overall quality (skill-quality-reviewer). |
Source Currency Auditor
Purpose
A skill can be perfectly written and quietly wrong because the external
world moved under it — a cited price changed, a linked doc 404s, a
research finding was superseded, a tool version is three majors behind.
None of it shows in the code; the library just slowly decays against
reality. This skill runs the broad currency sweep: it inventories every
external-source citation across the library, checks each against a
known-good source list and a staleness threshold, and flags what's stale,
broken, or superseded — for re-verification, prioritized by how
load-bearing and how volatile the claim is. It is a periodic gate on
external-truth decay, complementary to framework-edition-tracker's
narrow focus on standard editions. It flags; it does not verify or fix —
because confirming currency requires checking the live source, not
trusting memory.
Use When
- Use when: auditing whether the external sources the library cites are
still current, across all skills.
- Use when: setting or applying a citation-staleness threshold (flag
citations older than N months for re-verification).
- Use when: periodically sweeping for stale, broken, or superseded
external references as a maintenance gate.
- Use when: a skill's claim depends on a volatile external fact (price,
model id, tool version, limit) that may have drifted.
- Do NOT use when: the task is tracking specific FRAMEWORK editions and
producing deltas — that is
framework-edition-tracker (narrow, edition-
specific); this is the broad age/staleness sweep.
- Do NOT use when: the task is proposing the specific mapping edits from a
delta — that is
framework-mapping-refresher.
- Do NOT use when: the task is reviewing a skill's overall QUALITY
(triggers, sections, evals) — that is
skill-quality-reviewer; this
audits only citation currency.
Inputs to Inspect
- The external-source citations across skills: standards editions,
external doc/URLs, research reports, tool/library versions, prices and
limits, and any dated factual claim — and where each is cited.
- The known-good source list: the current/authoritative source for each
cited fact, where one exists, to compare against.
- The staleness threshold(s): the age (N months) beyond which a citation
needs re-verification — stricter for volatile facts than stable ones.
- Volatility of each fact: how fast it goes stale (prices/model-ids/tool
versions fast; foundational research slow) — sets the effective
threshold.
- Load-bearing-ness: whether the citation underpins a core claim or a
footnote — sets priority.
Workflow
- Inventory external-source citations. Sweep the library for every
claim resting on an external source: standard editions, URLs, research,
versions, prices/limits, dated facts. Capture what's cited, where, and
its stated date/version if any. An un-inventoried citation can't be
audited.
- Set thresholds by volatility. One flat "N months" is too blunt:
prices and model ids stale in weeks, tool versions in months, and
foundational research in years. Set an effective threshold per fact
class, stricter where the world moves faster.
- Check each citation against the known-good list. Classify: current,
STALE (older than its threshold — needs re-verification), SUPERSEDED (a
newer source exists), or BROKEN (dead link/removed source). Record the
reason per flag.
- Prioritize by load-bearing × volatility. A stale citation
underpinning a core recommendation outranks a stale aside; a volatile
fact outranks a stable one. Order the flags so re-verification effort
goes where wrongness costs most.
- Flag for re-verification — verify nothing yourself. Currency can
only be confirmed against the live source, not asserted from memory.
Produce the list of citations needing re-check, each with reason and
priority; the actual re-verification (and any fix) is a separate,
human/owner step. Route framework-edition specifics to
framework-edition-tracker for the deep dive.
- Recommend a cadence. This is a periodic gate — state how often to
run it (and tighter for the volatile-fact skills), so currency stays a
habit, not a one-off.
- Deliver the currency audit in the Output Format: inventory, flags
with reason and priority, and the cadence — changing nothing.
The citation-class thresholds, the flag taxonomy (stale/superseded/
broken), and the load-bearing × volatility priority matrix:
references/currency-audit-sheet.md.
Output Format
SOURCE CURRENCY AUDIT
Inventory: <cited fact | source | where cited | stated date/version | class>
Thresholds: per fact-class (volatile: weeks/months; stable: years)
Flags:
<citation>: STALE (age>threshold) | SUPERSEDED (newer source) | BROKEN (dead) — reason
Priority: load-bearing × volatility — highest first
Re-verify: list for HUMAN/owner re-check (this skill verifies nothing);
framework-edition specifics → framework-edition-tracker
Cadence: recommended re-run interval (tighter for volatile-fact skills)
Changes: NONE (audit only)
Boundaries: editions/deltas → framework-edition-tracker; propose edits →
framework-mapping-refresher; skill quality → skill-quality-reviewer
Validation Checklist
Gotchas
- Confirming a source is current requires checking the LIVE source;
asserting "still current" from memory is how the auditor becomes the
stale claim. This skill flags for re-verification — it does not certify
currency itself.
- A flat staleness threshold mis-serves both ends: it nags about
foundational research that's fine for years and misses a price that went
stale last week. Thresholds must track volatility.
- Broken links are the easiest currency failure to catch and the most
embarrassing to leave — a dead source is a claim with no backing. Flag
them first.
- Superseded is worse than merely old: a citation to a withdrawn or
replaced source actively misinforms, where a still-valid old source just
ages. Distinguish the two.
- Not all stale citations matter equally — prioritize by what the claim
holds up. A stale footnote is noise; a stale core recommendation is a
real defect.
- This is the BROAD sweep. When a flag is specifically a framework edition
change, hand the deep dive to
framework-edition-tracker; don't
reproduce its delta work here.
- Auditing currency is not reviewing quality. If you're assessing
triggers, sections, or eval integrity, that's
skill-quality-reviewer.
Stop Conditions
- The task is tracking specific framework editions / producing deltas →
route to
framework-edition-tracker.
- The task is proposing the specific mapping edits, or reviewing a skill's
overall quality → route to
framework-mapping-refresher or
skill-quality-reviewer.
- A citation's currency cannot be confirmed without checking the live
source → flag it for human re-verification with its reason; do NOT
certify it current (or stale) from memory.
- A flagged citation is load-bearing and its staleness would materially
mislead users → raise its priority and recommend prompt re-verification
rather than leaving it in a long backlog.
Supporting Files
- references/currency-audit-sheet.md
— the citation-class thresholds, the flag taxonomy (stale/superseded/
broken), and the load-bearing × volatility priority matrix.
evals/evals.json — behavior cases including volatility-tuned
thresholds, the flag-don't-verify discipline, and the broken-link catch.
evals/trigger-evals.json — discrimination against framework-edition-tracker
(editions vs broad staleness), framework-mapping-refresher (audit vs
propose), and skill-quality-reviewer (currency vs quality).