anonymized-pattern-extraction
/anonymized-pattern-extraction
Wyciąga ogólne wzorce z prywatnego repozytorium bez szczegółów identyfikujących, z przykładami syntetycznymi i kontrolą przed udostępnieniem.
- Autor
- tikoci ↗
- Repozytorium
- routeros-muse ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 04.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 04.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Wyciąga ogólne wzorce z prywatnego repozytorium bez szczegółów identyfikujących, z przykładami syntetycznymi i kontrolą przed udostępnieniem.
Oryginalny opis (EN)
Mine a private repo for generic patterns with zero private specifics: shape-not-detail extraction, labeled claims, synthetic examples, scrub-gated handoff. Use when: learning from a private workspace, writing up lessons from a network you cannot show, reviewing whether a write-up leaks, or asking what may travel from private to shared.
Jak działa
- Part 1 — The checklist (nothing private travels)
- Part 2 — Skill doc
Podgląd SKILL.md
---
name: anonymized-pattern-extraction
description: "Mine a private repo for generic patterns with zero private specifics: shape-not-detail extraction, labeled claims, synthetic examples, scrub-gated handoff. Use when: learning from a private workspace, writing up lessons from a network you cannot show, reviewing whether a write-up leaks, or asking what may travel from private to shared."
topics: [privacy, patterns, extraction, documentation]
tested: none
---
# `anonymized-pattern-extraction` skill (scope → extract → label → gate) [template v1 draft]
Promoted from an extraction assignment
(pattern only, shape-not-detail), its executed precedent (a
brief → feedback-doc pipeline that actually ran), the
carry-across rule, and the standing privacy lines.
Pairs with pre-publish-scrub: extraction
produces the doc, pre-publish-scrub gates its publication. Core rule: carry
across MEANING — never private inventories, addresses,
credentials, topology, code, or deployment-specific
assumptions. No secrets below; this skill shows shapes of
shapes.
## Part 1 — The checklist (nothing private travels)
Run top to bottom. Extraction is done when the write-up teaches
everything the private repo taught you and contains nothing an
owner would flinch to see quoted.
1. Scope the extraction (§0): assignment shape, what is wanted,
what is forbidden.
2. Extract shapes, not details (§1): patterns, lessons, failure
modes — every private value replaced by its shape.
3. Label every claim (§2): VERIFIED / JOURNALED / OPINION +
house style, contradictions named.
4. Gate and hand off (§3): synthetic examples, self-review,
pre-publish scrub, delivery.
### 0. Scope (the assignment, quoted)
The brief's assignment, verbatim in spirit (collected: an
extraction brief): mine the private repo for
GENERIC ideas — how tools were wired together, the auth scheme
it used, what broke, what worked, what you would reuse — and
write the PATTERNS AND LESSONS into the shared folder.
Patterns only: no private code, names, addresses, credentials,
or topology specifics. If a lesson depends on a private
detail, describe its SHAPE, not the detail.
- Wanted output (collected: brief "What to produce"):
architecture shape, auth scheme, failure modes, reuse-or-
avoid judgments; feedback on the live design from the
private vantage point; story-vs-results flags; test results
with exact commands + full outputs, environment labeled.
- Forbidden cargo (collected: the carry-across rule +
the agent-feedback privacy line): names, places, addresses,
serials, credentials, topology, private code, deployment-
specific assumptions. Public repo URLs and skill names only.
- The executed precedent (collected: the brief ran): an
operator feedback doc EXISTS — a private-network
operator read the brief, mined their repo, and delivered
patterns + feedback + flags + tests with the labels below.
Extraction is a proven pipeline, not a proposal.
- Scope edge: a lesson that cannot survive without its
private detail does not travel — it stays as a pointer
("ask the operator about the X-shaped failure") or it
waits for a synthetic reproduction.
### 1. Extract (shape-not-detail, every line)
- The transform: private value → shape. An address becomes
"a single A record, TTL 60" or "an RFC1918 range"; a
credential becomes "a Bearer [REDACTED] the relay"; a topology
becomes "two sites, one authenticating against the other";
a password becomes ABSENT (not a shape — the sticker red
line: identifiers extracted, the secret masked local-only,
never logged, never in a ledger — collected: the auth
red line).
- Patterns worth extracting (collected: brief + concept
review): wiring shapes (how tools composed), auth scheme
shape (what proved what, where secrets lived), failure
modes (what broke, what the symptom was, what fixed it),
reuse judgments (what to copy, what to avoid, and why),
measurement shapes (timeouts, counts, sizes — numbers
travel, identities do not).
- Commit-history mining (collected: the concept review's
inputs): decisions files + backlog + whiteboard history
carry ownership, identity, capture provenance, and recorded
evolution — mine the EVOLUTION (what changed and why), not
the contents. Rough-count precedent: "~80 field-work, ~50
config, ~20 dashboard, ~20 tools over 150 commits" teaches
effort shape with zero private content (collected: an
earlier commit census).
- Live by the carry-across sentence (collected: concept
review): carry across their MEANING, not private
inventories, addresses, credentials, topology, code, or
deployment-specific assumptions. When unsure, the meaning
without the noun: "the tunnel token leaked through an
unredacted env line" teaches; the token itself must never
appear.
### 2. Label (every claim carries status + evidence)
Claim labels from the brief (collected: the brief's "What to
produce" + environment sections):
- [VERIFIED] = you ran/checked it in session, source named.
- [JOURNALED] = repo history or another agent's record says
it happened (cite the record, not your memory of it).
- [OPINION] = your recommendation (reasons shown, not
asserted).
- [APPLICANT]/[DEVELOPER] = the human stated it (source
linked, date noted). [PENDING] = still to confirm.
[ASK MUSE] = only the named environment can check it.
House style, because someone will ask "did you test it"
(collected: brief closing): every claim carries its evidence
AND its status; never let PENDING read as DONE; if live
evidence contradicts a doc (including the brief itself), say
so and name the doc/line. The grounds-file pattern is the
same discipline at scale: every application claim traces to
a line with a status tag (collected: grounds and references,
[VERIFIED]/[APPLICANT]/[PENDING] tags).
- Evidence-ledger compatibility (evidence-ledger): observed/collected/
inferred/pending labels map onto these 1:1 — use whichever
vocabulary the consuming doc uses, but never an unlabeled
claim.
- Contradictions are findings, not embarrassments: when
environments disagree (sandbox DNS vs host dig, TCG vs HVF
timing), record WHICH environment each result came from —
reach differences are data, not contradictions (collected:
the brief's environment notes).
### 3. Gate (synthetic examples, scrub, handoff)
- Examples are SYNTHETIC: "examples are synthetic; no private
locations or topology are reproduced" and the closing
attestation "no private network identities, addresses,
credentials, or topology are included" (collected: the
concept review's framing + closing). Write the example from the
shape, then re-read it asking "could the operator's
neighbor identify the network from this" — if yes, blur
further.
- Self-review before handoff: grep the draft for the forbidden
cargo (§0); read it as the operator (flinch test); confirm
every private-dependent lesson survived as shape-only or
was cut to a pointer.
- pre-publish-scrub gates publication: runtime-state sweep,
sensitivity tiers, zero-findings gate, clean-home staging,
pair review. The extraction draft enters pre-publish-scrub
like any other publish candidate — extraction care does not
waive the scrub.
- Handoff shape (collected: brief delivery): one doc, sections
for patterns / design feedback / story-vs-results flags /
test results, changes listed at top if revising. Point at
state (paths, commits, record IDs); do not paste copies of
private sources — not even "just for context".
## Part 2 — Skill doc
### Trigger
Load when: learning from a private workspace; writing up
lessons from a network you cannot show; reviewing whether a
write-up leaks; "what may travel from private to shared"
questions; scoping someone else's extraction assignment.
(Entry-skill + capability-probe pattern.)
### Prereqs
- Read access to the private source + the operator's consent
for extraction (not just access — consent to PUBLISH
patterns from it).
- The consuming doc's claim vocabulary chosen (brief labels
vs evidence-ledger labels).
- pre-publish-scrub available for the publication gate.
### Access
- Reads of the private source follow its own access rules;
extraction adds no new read rights.
- The draft lives in the SHARED folder from the first line —
never draft extraction inside the private repo and "move it
later" (the move is where leaks hide).
- Operator review before any publish step: the operator who
owns the private source reads the draft, not just the
scrub report.
- No private source is pasted, quoted at length, or attached
to the draft "for reference" — shapes only, always.
### Envs
| Name | Use |
|---|---|
| (none required) | Extraction is reading + writing; no service env |
Private-repo credentials (if any) stay in their existing
store and take no part in the draft. No env carries private
content toward the shared folder.
### Outputs
- Scoped assignment (wanted + forbidden, §0).
- Pattern write-up: wiring/auth/failure/reuse shapes.
- Labeled claims (every claim with status + evidence).
- Synthetic examples + flinch-test pass note.
- pre-publish-scrub record + handoff doc.
### Verification
- Forbidden-cargo grep over the draft: zero private code,
names, addresses, credentials, topology, serials.
- Every claim labeled; every PENDING visibly pending; every
contradiction naming doc/line.
- Every example synthetic (attestation line present).
- Operator read confirmed (not assumed from access).
- pre-publish-scrub gate green before any publish step.
- Claim audit: each extraction claim carries source + status
or an honest pending label.
### Recovery
- Suspected leak in the draft → delete the value, blur to
shape, re-run the flinch test; if already shared, rotate
if live and record as an anomaly (evidence-ledger
precedent).
- Private-dependent lesson that will not blur → cut to a
pointer ("ask the operator") or commission a synthetic
reproduction; never publish the detail "just this once".
- Operator flinch at review → the operator is right: blur
further or cut, no argument from the extractor.
- Label dispute → the weaker label wins (VERIFIED needs a
run you can point at; otherwise it is JOURNALED).
### Limits
- Extraction is lossy by design: some lessons cannot travel
— the skill names the loss (pointer), never smuggles the
detail.
- Synthetic examples teach shapes, not instances: a reader
who needs the instance needs operator access, not a better
example.
- The brief → feedback precedent ran ONCE, in one direction
(workstation → shared folder) — other environment pairs re-prove the
pipeline, they do not inherit it.
- Operator consent is per-extraction: access last month is
not consent this month.
Wersje
W bazie nie ma innych skilli o nazwie anonymized-pattern-extraction.
Migawka bazy z 13.11.2026 · wersja paczki 0d15a390ef20 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100pre-publish-scrub
Czyści przestrzeń roboczą przed udostępnieniem: stan wykonania, identyfikatory, EXIF zdjęć i syntetyczne dane testowe.