goal-loop
/goal-loop
Prowadzi zadanie w pętli celów: spisuje kontrakt, wykonuje pracę, weryfikuje ją niezależnym agentem i powtarza do zaliczenia.
- Autor
- meshachjackson ↗
- Repozytorium
- agent-skills ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 07.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 07.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Prowadzi zadanie w pętli celów: spisuje kontrakt, wykonuje pracę, weryfikuje ją niezależnym agentem i powtarza do zaliczenia.
Wyzwalacze z opisu
Oryginalny opis (EN)
Run a task as a goal loop. Write a contract (goals, exit criteria, checks, budget), work, verify with an independent agent, repeat on FAIL until PASS or budget. Use for "goal loop", "verify until done", or exit criteria plus a verifier.
Jak działa
- 1. Contract (before any work)
- 2. Review the contract
- 3. Work
- 4. Verify
- 5. Decide
- 6. Report
- Hard rules
Podgląd SKILL.md
--- name: goal-loop description: Run a task as a goal loop. Write a contract (goals, exit criteria, checks, budget), work, verify with an independent agent, repeat on FAIL until PASS or budget. Use for "goal loop", "verify until done", or exit criteria plus a verifier. version: 1.1.0 --- # goal-loop Requires the `loop-verifier` agent (`agents/loop-verifier.md`) installed in `~/.claude/agents/`. Without it, stop and say so. You do the work; the `loop-verifier` agent grades it. Never grade your own work. Use for goal loops, exit criteria, verifiers, or "keep going until it passes", when "done" is easy to fake. Skip small one-step tasks. ## 1. Contract (before any work) Write it to the repo's plan location (e.g. `docs/plans/<yyyy-mm-dd>-<slug>.md`), else the scratchpad. Parts: - **Request.** The user's words, verbatim. Criteria are judged against this, not your reading of it. - **Goals.** Three at most, one or two sentences. - **Exit criteria.** Numbered. Each names a count, file, command or field (no "good", "clear", "thorough") and has a `Check:` line: the command, or the file and what to look for. If you cannot write one, rewrite or drop the criterion. - **Tests.** Commands that must pass. For analysis tasks, every `file:line` claim must be real. - **Budget.** Max iterations, default 4 (one work pass plus one verify). - **Rules.** What is off limits (writes, pushes, deploys, new dependencies, paid or rate-limited calls), from the repo's `CLAUDE.md` and the user's prompt. Copy the repo's `CLAUDE.md` rules in; do not just point at the file. **Check-writing pitfalls.** These caused most contract review failures: - Shell quoting: `\$` inside double quotes. Use single quotes, or test the command once. - A grep pattern that matches its own source (the contract, the check line, the script). Anchor it or exclude the file. - Pipe exit codes: `cmd | grep x` hides the failure of `cmd`. Use `set -o pipefail` or check each step. - Diffing against a moving `origin/main`. Name a fixed base, such as `origin/main...HEAD` after a fetch, or a commit SHA. - Missing repo rules: read the repo's `CLAUDE.md` and copy its constraints into Rules. After the first verify, changing the contract needs the user's OK. If a criterion is wrong or impossible, stop and ask. ## 2. Review the contract Spawn `loop-verifier` with `mode: review contract` and the contract path only. It uses a fixed checklist. Round 2 passes `recheck only: <items flagged>`; the verifier then re-checks only those and adds nothing new. Cap: at most 2 review rounds. A review round is not an iteration. If round 2 still fails, stop reviewing, proceed to work, and write under Rules: `Contract review capped at 2 rounds; the final work verify is the gate. Open review items: <list>.` Ask the user only if a criterion is impossible or the Request is unclear. If the reply lacks `Scope: contract`, the agent definition is stale: say so and stop. Then show the user the contract as a short list and start, unless they asked to approve it first or it needs a risky action. ## 3. Work One pass against the contract; save output where it says. ## 4. Verify Spawn `loop-verifier` with only the contract path, the output path, and from iteration 2 on `recheck only: <numbers of the criteria that failed>`. Rechecks can miss regressions, so PASS is final only after one pass over all criteria. If `loop-verifier` is unavailable, say so and stop. ## 5. Decide - **PASS on a full pass:** report. - **PASS on a recheck round:** run one full pass. It is not a new iteration unless it fails. - **FAIL, budget left:** fix only the fix-list items, then verify with a recheck list. - **FAIL, budget spent:** stop and report what still fails. - **Same criterion fails twice for the same reason:** stop and ask the user. ## 6. Report Plain language, headline first: PASS in N iterations, or FAIL with what is left. Then, skipping empty ones: What I did; Verifier result; Blocked on me; Blocked on you (with the reason); Ready. Name the contract path. End with `Cost:` iterations used, and verifier calls split into review, recheck and full. ## Hard rules - The contract exists before the first work pass. - The verifier sees the contract and output, not your notes. - An uncheckable criterion counts as failed. - Never edit the contract to make a failing check pass. - Repo rules win. This skill grants no permission to push, deploy, write protected tables or add dependencies. - Stay inside the budget; more iterations need the user's OK.
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · meshachjackson/agent-skills | 90 | MIT |
| davidondrej/skills | 100 | MIT |
| Idfoxes/keelson | 100 | Apache-2.0 |
Migawka bazy z 13.11.2026 · wersja paczki 966cbb44e554 (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
- 100loop-goal
Ustala zasady dyscypliny dla długotrwałych zadań agenta: pętli cyklicznych i celów wykonywanych do spełnienia kryterium ukończenia, gdy zadanie powtarza się w odstępach lub działa autonomicznie. - 100goal-verifier
Weryfikuje ukończenie zadania względem założonych celów, zanim praca zostanie uznana za skończoną. - 100simmer
Prowadzi pętlę celu, w której Codex wykonuje kolejne przebiegi, a Claude sprawdza każdy wynik aż do zaliczenia. - 100codex-verify
Weryfikuje plan lub dokument z Codex jako niezależnym recenzentem i wydaje werdykt PASS lub FAIL. - 100loop
Prowadzi ograniczone powtarzanie na pierwszym planie w bieżącej sesji: tworzy kontrakt pętli, wykonuje akcję z weryfikatorem do zadeklarowanego limitu prób i zapisuje dowody w .planning/loops/. - 85orchestration-prompt
Generuje gotowy do uruchomienia prompt orkiestracji z weryfikowalnym celem, ograniczoną pętlą i niezależnym sprawdzeniem twórca-weryfikator, odrzucając niejasne cele.