Przejdź do treści
Biblioteka skilli
AI i agenci · Nauka i wyjaśnianie

wdym

/wdym

Rozszyfrowuje dla nadzorującego użytkownika wyniki pracy agenta w sesji: wnioski, surowe logi, diffy, stack trace'y, wyniki zapytań i raporty subagentów, opierając się na dowodach z sesji.

Licencja
Apache-2.0
Ostatnia zmiana repozytorium
21.10.2026

Ocena i bezpieczeństwo

100/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść35 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
2
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 21.10.2026, bez odwołań do wycofanych modeli i API.

Licencja

Apache-2.0, otwarta: można używać i zmieniać, z zachowaniem licencji i informacji o zmianach.

Co robi i kiedy

Rozszyfrowuje dla nadzorującego użytkownika wyniki pracy agenta w sesji: wnioski, surowe logi, diffy, stack trace'y, wyniki zapytań i raporty subagentów, opierając się na dowodach z sesji.

Wyzwalacze z opisu

wdymwhat do you meanI donwhat just happenedwhat does this output meanwalk me through what you didwhat is Xwhy this way
Oryginalny opis (EN)

Decode this session's agent output for the supervising user — prose conclusions, raw tool output (logs, diffs, stack traces, query results), or subagent reports — grounded in the session's own evidence. Trigger on self-directed comprehension or verification asks with no named audience: 'wdym', 'what do you mean', 'I don't follow', 'I don't understand', 'what just happened', 'what does this output mean', 'walk me through what you did', mid-task 'what is X', 'why this way', 'how did you verify', 'prove it'. Also 'wdym debrief' for an end-of-task comprehension pass. Do NOT trigger when an audienc

Jak działa

  1. Step 0: Load the profile
  2. Calibration (generic defaults; the profile overrides)
  3. The ladder
  4. Format escalation
  5. Verification posture
  6. Learning loop
  7. Promotion
  8. Debrief mode
  9. Boundaries
Sekcje z nagłówków SKILL.md, w oryginale · 94 linii, 646 słów
Podgląd SKILL.md
---
name: wdym
description: "Decode this session's agent output for the supervising user — prose conclusions, raw tool output (logs, diffs, stack traces, query results), or subagent reports — grounded in the session's own evidence. Trigger on self-directed comprehension or verification asks with no named audience: 'wdym', 'what do you mean', 'I don't follow', 'I don't understand', 'what just happened', 'what does this output mean', 'walk me through what you did', mid-task 'what is X', 'why this way', 'how did you verify', 'prove it'. Also 'wdym debrief' for an end-of-task comprehension pass. Do NOT trigger when an audience is named ('explain to my manager', 'ELI5 for my wife' — that is the eli5 skill) or for pure visual requests (show-me)."
user-invocable: true
---

# WDYM — What Do You Mean

Re-explain a stretch of this session's agent work to the person supervising it. The source
material is the session itself: explain what actually happened here, never the general case.

## Step 0: Load the profile

```bash
source "${CLAUDE_PLUGIN_ROOT}/skills/common/get_kb_path.sh" && KB_PATH=$(get_kb_path)
cat "$KB_PATH/rules/comprehension-profile.md"
```

The profile carries the person's calibration that is not already in their always-on
instructions — register, analogy policy, native formats, daily-stack anchors, known confusion
clusters — and overrides the generic defaults below. If the file is missing, use the defaults
and offer once to compile a profile from the knowledge bank's reflections.

## Calibration (generic defaults; the profile overrides)

- Peer register for a working engineer. Never a child's register.
- Prose at "80% of the way to ASD-STE100": short sentences, active voice, one term per
  concept, one idea per sentence. Soften rather than follow the spec strictly — strict STE
  measurably drops facts.
- One brief analogy may open; the in-context example is mandatory: the actual file, log line,
  or command from this session. An analogy without the in-context example is a non-answer.
- Anything with layers, roles, or directions gets a 2-axis table or a decision tree, not prose.
- Close substantial explanations with a trigger list: "you'll know X applies when …".

## The ladder

Follow-ups escalate predictably. Answer the rungs that apply, and never leave the NEXT rung
unaddressed — it is the question you will be asked anyway:

1. What changed?
2. Where does it come from? (file, config, upstream system)
3. What is X? (term, acronym — define once, use it consistently after)
4. Why this way? (the alternative considered, and why not)
5. Prove it. (the verification evidence)

## Format escalation

Start at rung 1; climb only while the point is not landing:

1. **Prose** — 80%-STE, calibrated as above.
2. **Diagram** — when the subject has layers, roles, directions, or flow. After drawing,
   restate every edge as a checkable sentence; a missing arrow silently claims "no dependency".
3. **Interactive HTML** — one focused page, only when a static diagram cannot carry it (state
   spaces, before/after comparisons, things worth poking at). Reuse the show-me grammar.

Explainer videos are out of scope.

## Verification posture

Clarity must not outrun truth: a wrong claim in a clean diagram is MORE convincing, not less.

- Every claim about what happened cites this session's evidence: `file:line`, the actual log
  line, the command and its output.
- Mark **verified** (evidence is in the session) vs **inferred** (reasoning, not yet checked)
  explicitly.
- If a verification gap exists, say so plainly — the "prove it" rung exists because gaps hide.

## Learning loop

When the person's confusion is new, or a known cluster repeats, append one dated line to the
profile's `## Inbox` section:

```
- YYYY-MM-DD: confused by <X> in <context>; <what explanation landed>
```

On "recompile my profile": fold the Inbox and recent reflections into the profile body, then
empty the Inbox.

## Promotion

After a substantial explanation that resolved real confusion, offer once — never auto-write —
to promote it to a knowledge-bank concept doc, preserving the analogy under a
`## The Analogy (from the session's ELI5)` heading. Use the kb-ingest workflow.

## Debrief mode

`wdym debrief` (or asked at task end): walk the ladder over the whole task — what changed,
where it came from, new terms, why this shape, and the verification evidence — pre-answering
the questions the person would otherwise ask one by one. Close with next steps: every pending
item, one line each, naming who holds it, or `none`.

## Boundaries

- A named audience ("explain this to my manager") → the eli5 skill, not this one.
- A standalone visual with no comprehension question → show-me.
- "How did we handle X before" / service documentation → knowledge-bank-lookup.
- Deliberate multi-session study of a topic → teach.
Pliki w paczce (2)
  • CHANGELOG.md 2 KB
  • SKILL.md 5 KB

Wersje

W bazie nie ma innych skilli o nazwie wdym.

Migawka bazy z 13.11.2026 · wersja paczki 07a8a5e1f06d (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

  • 100analyzing-cc-sessions
    Analizuje surowe logi sesji Claude Code: parsuje transkrypty JSONL, bada zachowanie agentów, wyciąga metryki i debuguje wywołania subagentów, także przy budowie narzędzi do analizy sesji.
  • 100agent-session-forensics
    Diagnozuje historię sesji agenta: przerwane pętle narzędzi, brakujące wyniki narzędzi, wąskie gardła czasowe i korelację śladów subagentów.
  • 100claude-session-parsing
    Znajduje i streszcza transkrypcje sesji Claude Code w plikach jsonl, w tym subagentów i wyników narzędzi, z kosztem, czasem i wywołaniami.
  • 100comprehension-gate
    Sprawdza, czy użytkownik rozumie zmianę zbudowaną przez agenta, zanim ją zaakceptuje, po testach i przy przekazaniu pracy.
  • 100analyzing-codex-sessions
    Analizuje surowe logi sesji Codex CLI: parsuje pliki rollout JSONL, bada zachowanie subagentów Codex, wyciąga metryki wątków i debuguje, dlaczego agent zawiódł.
  • 100skill-debrief
    Robi przegląd skilla po użyciu: zbiera uwagi z sesji i wprowadza poprawki do SKILL.md, gdy nie uruchomił się tak, jak powinien.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.