Przejdź do treści
Biblioteka skilli
AI i agenci · Planowanie

deep-plan

/deep-plan

Dzieli zadanie na etapy: najpierw plan zapisany w pliku, potem czyszczenie kontekstu i wykonanie z nowej sesji.

Repozytorium
scripta ↗
Licencja
Apache-2.0
Ostatnia zmiana repozytorium
13.11.2026

Ocena i bezpieczeństwo

90/ 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.
Treść25 / 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
1
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 13.11.2026, 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

Dzieli zadanie na etapy: najpierw plan zapisany w pliku, potem czyszczenie kontekstu i wykonanie z nowej sesji.

Wyzwalacze z opisu

plan this properlyletexecute PLAN-<anything>.mds execute phase), orm running out of context
Oryginalny opis (EN)

The enforced plan → clear → execute flow — think with full context, write the plan to a file, then execute from a fresh session that reads only the plan. Use when the user says "/deep-plan", "plan this properly", "let's not burn context", "execute PLAN-<anything>.md" (that phrase IS this skill's execute phase), or "/deep-plan handoff" / "capture state, I'm running out of context" mid-task. The token-saving best practice as one command instead of a habit everyone must remember.

Jak działa

  1. Phase 1 — plan (this session, context is disposable)
  2. Phase 1b — mid-task handoff (entering with the context already fat)
  3. Phase 2 — the handoff (the enforced part)
  4. Phase 3 — execute (fresh session, context is precious)
  5. Guardrails
Sekcje z nagłówków SKILL.md, w oryginale · 87 linii, 836 słów
Podgląd SKILL.md
---
name: deep-plan
description: The enforced plan → clear → execute flow — think with full context, write the plan to a file, then execute from a fresh session that reads only the plan. Use when the user says "/deep-plan", "plan this properly", "let's not burn context", "execute PLAN-<anything>.md" (that phrase IS this skill's execute phase), or "/deep-plan handoff" / "capture state, I'm running out of context" mid-task. The token-saving best practice as one command instead of a habit everyone must remember.
---

# deep-plan

One loop: **explore fat → plan to a file → clear → execute thin.** The context spent
understanding a problem is dead weight during execution; this skill makes dropping it
the default instead of a discipline. The fix is a flow, not a reminder.

Do NOT use the built-in plan mode (EnterPlanMode/ExitPlanMode) inside this flow: its
approve-then-proceed step continues in the same fat session, which is exactly what
Phase 2 exists to prevent. This skill replaces it here, not wraps it.

## Phase 1 — plan (this session, context is disposable)

1. Explore as widely as the task needs — read code, search, ask the user. This session's
   context is about to be thrown away, so spending it is free.
2. Write the plan to `PLAN-<slug>.md` in the current repo root. Immediately append
   `PLAN-*.md` to that repo's `.git/info/exclude` (NOT `.gitignore` — no PR, works in
   any checkout, keeps the file out of `git add -A` sweeps and out of the handbook's
   freshness lint). Work spanning multiple repos: ONE plan file, in the repo where
   execution starts; every file reference in it is an absolute path.
3. The plan must be executable by a session that has seen NOTHING except this file.
   Hard test for every line: "does this sentence survive without the conversation?"
   It starts with this header, verbatim, so the flow survives even if this skill
   doesn't load in the fresh session:

   > This is a deep-plan execution file. Read only the files named here — do not
   > re-explore. Work the steps in order, checking each off in this file as it lands.
   > A fact missing here is a plan bug: ask, then append the answer here.

   Then:
   - **Goal** — one sentence, plus how to verify it's done (command, test, URL).
   - **Start in** — the absolute directory the execute session should run from.
   - **Context a stranger needs** — files involved (absolute paths + line refs),
     constraints/conventions found while exploring, decisions already made with the
     user. Facts, not narration of the exploration.
   - **Steps** — numbered checkboxes, each independently checkable, smallest that works.
   - **Not doing** — what was considered and excluded, so the executor doesn't
     re-litigate it.
4. Show the user the plan. Iterate until they approve. No approval → no phase 2.

## Phase 1b — mid-task handoff (entering with the context already fat)

"/deep-plan handoff", "capture state", or the user is deep in a task at high context:
same as Phase 1 but the exploration already happened — capture instead of explore.
The plan file additionally records **current state** (what already works), **changes
made so far** (files touched, committed or not), and **next steps** as the Steps list.
Then hand off exactly as Phase 2. This is the "save my session before it drowns" path.

## Phase 2 — the handoff (the enforced part)

On approval, do NOT keep working. Give the user this handoff (with the actual
filename, not the placeholder) and stop:

> Plan saved to `PLAN-<slug>.md`. Run `/clear`, then say: **execute PLAN-<slug>.md**.

That stop is the skill's entire value: continuing in the fat session is what burns the
tokens. Do not "just do step 1 quickly" in the planning session. If the user says
"execute it" / "go ahead" WITHOUT having cleared — this session still contains the
Phase 1 exploration — refuse once and repeat the handoff line; executing fat because
the user forgot defeats the flow silently, which is worse than a reminder.

## Phase 3 — execute (fresh session, context is precious)

When invoked as "execute PLAN-<file>.md":

1. Read the plan file, `cd` to its **Start in** directory. Read ONLY the files it
   names. Do not re-explore; the plan is the exploration's output. Missing fact in
   the plan → that's a plan bug: ask the user, then append the answer to the plan
   file so the next session has it too.
2. Work the steps in order, checking each off in the plan file (`- [x]`) as it lands —
   the file is the progress tracker, so an interrupted run resumes by reading it.
3. Done = the Goal's verify line passes. Then delete the plan file (or offer to), and
   if real decisions or durable facts surfaced, offer `/document` (facts → handbook)
   or the daily-digest ADR path (decisions) — don't let them die with the session.

## Guardrails

- Plan files are scaffolding, never committed (`.git/info/exclude` in Phase 1 enforces
  it). Durable facts belong in the handbook via `/document`, decisions in `decisions/`
  — not in plan files.
- Self-invoked on a size judgment (the user didn't ask): confirm in one line first —
  "this looks big; plan-then-clear, or straight through?" — the ceremony is only worth
  it when the user agrees the task is big.
- User-invoked: Phase 1 always ends at the handoff, even when the task turned out small.
- Phase 3 that finds the plan wrong stops and says so; it never silently re-plans.
  Re-planning is Phase 1, and it's the user's call to restart it.

Wersje

RepozytoriumWynikLicencja
Ta strona · team-blue-Hub/scripta90Apache-2.0
oguzsabi/deep-skills100MIT
piercelamb/deep-plan100MIT
sungmin-cho/deep-work81MIT

Migawka bazy z 13.11.2026 · wersja paczki 32aa521b9087 (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

  • 100plan-execute
    Koordynuje wykonanie zatwierdzonego pliku planu między Claude a Codexem, z przeglądem i poprawkami do progu jakości.
  • 90plan-loop
    Planuje pracę z użytkownikiem w pliku PLAN.md, a potem implementuje ją krok po kroku, z /clear po zapełnieniu kontekstu, także przy wznawianiu pracy.
  • 100handoff-execute
    Wykonuje brief przekazania z wymuszoną dyscypliną czytania źródeł: wczytuje wymagane źródła przed planem, cytuje je w krokach, uruchamia weryfikację wyniku i zapisuje retrospektywę.
  • 100compact-pro
    Zapisuje stan sesji Claude Code do pliku przekazania i kopii transkryptu, a potem czyści sesję w bezpiecznym miejscu.
  • 100context-manager
    Zarządza oknem kontekstu i przekazaniami sesji w długich budowach i wieloagentowych: obserwuje zużycie, dobiera strategię kompaktowania i pisze oraz waliduje ustrukturyzowany plik przekazania, by nowa sesja podjęła pracę bez strat.
  • 90plan-execute
    Stosuje ostatni zatwierdzony plan dokładnie tak, jak jest zapisany, i jawnie zgłasza decyzje wymagające oceny zamiast podejmować je po cichu.

Uwagi i zgłoszenia

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