align
/align
Pokazuje przed działaniem, jak rozumie zadanie, w postaci manifestu ujawniającego ukryte założenia, gdy prośba jest niejednoznaczna, niedookreślona lub zbyt ogólna, by zacząć pracę.
- Autor
- pandysp ↗
- Repozytorium
- claude-plugins ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 20.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 20.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Pokazuje przed działaniem, jak rozumie zadanie, w postaci manifestu ujawniającego ukryte założenia, gdy prośba jest niejednoznaczna, niedookreślona lub zbyt ogólna, by zacząć pracę.
Wyzwalacze z opisu
Oryginalny opis (EN)
Surface what I think we're doing before acting on it. Use when the user invokes the align skill, says "let's align first", "check in before you start", or "mirror what you're hearing". Also invoke before starting any task with real interpretive weight. An ambiguous request, a vague spec, a "build X" with the details unstated. Where drafting from the wrong premise would waste the work. Produces a structured manifest exposing every silent commitment so the user can correct course before any artifact is produced.
Jak działa
- The manifest
- After the manifest
- Common pitfalls
Podgląd SKILL.md
--- name: align description: Surface what I think we're doing before acting on it. Use when the user invokes the align skill, says "let's align first", "check in before you start", or "mirror what you're hearing". Also invoke before starting any task with real interpretive weight. An ambiguous request, a vague spec, a "build X" with the details unstated. Where drafting from the wrong premise would waste the work. Produces a structured manifest exposing every silent commitment so the user can correct course before any artifact is produced. --- # Align: share understanding before acting The most expensive failures aren't bad outputs. They're outputs built on a wrong premise. By the time the user sees prose, you've silently committed to: your reading of the request, your chosen approach, your interpretations of source, your register and format choices, and a hundred small fill-ins for things that weren't specified. Each commitment looks authoritative once written. This skill inverts that. Before producing any artifact, produce a **manifest**: a structured, scannable summary of every silent commitment. The user reacts. Drafting happens after. ## The manifest Output exactly five blocks. Target <60 seconds to read. If a block balloons, the source is thin or the scope is too broad. Surface that, don't pad. Tag every item: - `[grounded]`: directly stated in the prompt or source. - `[inferred from "X"]`: derived from grounded content; show the chain. - `[mine]`: autonomously introduced; no derivation from source. Tags answer different questions: grounded → confirm, inferred → correct the premise, mine → accept or override. ### Block 1: My read of the task Paraphrase the request. State the audience. State **why** this matters / what it's for. Note any timing constraints. Form details (length, voice, structure) belong in Block 4. ### Block 2: My approach Describe the method. Include scope: what to do AND what to explicitly NOT do. Surfaces silent strategy commitments. Asked for a draft, got a strategy doc; asked for a bug fix, got a refactor. ### Block 3: What I'm working with **Source coverage**: what's in the source, tagged. Format: ``` - Source fact `[grounded]` - Derivation from source `[inferred from "..."]` ``` ### Block 4: How I'd render it State the **shape** (length, structure, container) and a **sample** of the actual output. The sample carries register, voice, conventions. Don't list style attributes the sample reveals on its own. Catches silent voice drift. Format: ``` - Shape: [length, structure, container] - Sample: *"[an actual opening line of the output, in the intended voice]"* ``` ### Block 5: Blocking questions Only forks where any default would be unsafe. If there are none, say so. Don't pad. ## After the manifest Stop. Wait for the user to confirm, correct, or re-scope. Once confirmed, **re-read the manifest before drafting**. The committed items are what you're building on. If something needs to change during drafting, pause and surface it: *"I'd like to add X. Wasn't in the manifest. OK?"* Don't drift silently. ## Common pitfalls - **Manifest theater.** Clean manifest, then drift in the draft. Re-read before drafting; surface deviations. - **Padding to look thorough.** Inflating blocks to fill space. If approach is trivial, one bullet is enough. - **Hiding inventions in `[inferred]`.** If there's no derivation chain to point at, it's `[mine]`. - **Treating the manifest as a contract.** It's a checkpoint. Drift is allowed *when surfaced*, not when silently absorbed. - **Confusing align with clarify.** Align commits to specific reads; clarify asks where you're stuck. Different modes. - **Summarizing from memory instead of reading source.** Block 3 requires actually reading what's there. If source is too long, say so explicitly.
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · pandysp/claude-plugins | 100 | MIT |
| tony/skills | 100 | MIT |
| andriiklymiuk/corgi | 100 | MIT |
| hmmhmmhm/aligns.md | 100 | MIT |
| justinedevs/meta-architect | 100 | MIT |
| hmmhmmhm/aligns.md | 90 | MIT |
| 70-10/claude-code-plugins | 82 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki a805cde28188 (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
- 100interview-me
Ustala, czego użytkownik naprawdę chce, zadając pytania po jednym aż do około 95 proc. pewności co do intencji, gdy prośba jest niedookreślona. - 100clarify-question-loop
Zadaje od jednego do pięciu celowanych pytań, gdy prośba do meta-agenta jest zbyt niejednoznaczna, by bezpiecznie generować, pakować, publikować lub adaptować. - 100interview-me
Wydobywa prawdziwe potrzeby użytkownika przez wywiad z jednym pytaniem naraz do około 95 proc. pewności co do intencji, gdy prośba jest niedookreślona lub użytkownik prosi o wywiad. - 100interview-me
Wydobywa prawdziwe potrzeby użytkownika przez wywiad z jednym pytaniem naraz do około 95 proc. pewności co do intencji, gdy prośba jest niedookreślona lub użytkownik prosi o wywiad.