Przejdź do treści
Biblioteka skilli
AI i agenci · Planowanie

goal-harness

/goal-harness

Zamienia niejasne zadanie w cel z kryterium weryfikacji i pętlą recenzent-wykonawca, która kończy się po spełnieniu rubryki.

Repozytorium
design-agency ↗
Licencja
MIT
Ostatnia zmiana repozytorium
02.11.2026

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
3
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 02.11.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co robi i kiedy

Zamienia niejasne zadanie w cel z kryterium weryfikacji i pętlą recenzent-wykonawca, która kończy się po spełnieniu rubryki.

Wyzwalacze z opisu

t want, or when writing an initiative
Oryginalny opis (EN)

Turn a fuzzy task into a goal with five elements — a concrete outcome, a verification that proves it, constraints that fence the scope, an iteration policy, and error handling — plus a rubric, then run it as an executor-reviewer loop that stops only when the rubric passes or a stop condition fires. Use when a task will run longer than 30 minutes or across many iterations unattended, when using a /goal, Ralph-loop, or executor/reviewer pattern, when an agent keeps finishing early with output you don't want, or when writing an initiative's .goal.md.

Jak działa

  1. When to use
  2. Steps
  3. Output format
  4. Watch-outs
  5. See also
Sekcje z nagłówków SKILL.md, w oryginale · 53 linii, 540 słów
Podgląd SKILL.md
---
name: goal-harness
title: Structure a long-running task as a five-element goal
description: "Turn a fuzzy task into a goal with five elements — a concrete outcome, a verification that proves it, constraints that fence the scope, an iteration policy, and error handling — plus a rubric, then run it as an executor-reviewer loop that stops only when the rubric passes or a stop condition fires. Use when a task will run longer than 30 minutes or across many iterations unattended, when using a /goal, Ralph-loop, or executor/reviewer pattern, when an agent keeps finishing early with output you don't want, or when writing an initiative's .goal.md."
status: beta
narrative: 00_README/_frameworks/long-horizon-execution.md
when_to_load:
  - A task will run longer than 30 minutes or across many iterations without supervision
  - Using a /goal, Ralph-loop, or executor/reviewer agent pattern
  - An agent keeps "finishing early" with output you don't want
  - Writing or reviewing a <slug>.goal.md for an optimization initiative
applies-to: long-running autonomous tasks, optimization initiatives, /optimize-loop runs, scheduled jobs that must reach a real completion bar
inputs:
  - kind: goal
    where: "{00_Agentic-OS,01_Project-Hubs/[!_]*}/**/*.goal.md"
outputs:
  - kind: goal-block
    schema: "A .goal.md whose front-matter declares outcome, verification, constraints, iteration_policy, error_handling, and rubric (path to a .rubric.md)"
done_when:
  - "The goal declares all five elements: outcome, verification, constraints, iteration_policy, error_handling"
  - "The outcome is a concrete target state (a number, threshold, or observable), not an adjective"
  - "Verification names a check that proves done — a tool, a test, or a rubric pass — never self-report"
  - "Constraints name what must not change; error handling says when to pause and report"
  - "A reviewer scores every round against the rubric before the loop may stop"
lints:
  - id: goal-outcome-is-adjective
    severity: error
    where: "{00_Agentic-OS,01_Project-Hubs/[!_]*}/**/*.goal.md"
    pattern: "^outcome:\\s*[\"']?(make it )?(better|nicer|cleaner|good|great|improved)\\b"
    message: "Outcome must be a concrete target state with a number, threshold, or observable — not an adjective like 'better'."
  - id: goal-missing-verification
    severity: error
    where: "{00_Agentic-OS,01_Project-Hubs/[!_]*}/**/*.goal.md"
    pattern: "^outcome:\\s*\\S"
    must_also_match: "^verification:\\s*\\S"
    message: "Every goal with an outcome must declare verification — how we prove it is done (tool, test, or rubric), not self-report."
  - id: goal-missing-error-handling
    severity: warning
    where: "{00_Agentic-OS,01_Project-Hubs/[!_]*}/**/*.goal.md"
    pattern: "^outcome:\\s*\\S"
    must_also_match: "^error_handling:\\s*\\S"
    message: "Long-running goals should declare error_handling — when to pause and report instead of looping."
budget:
  context_tokens_max: 2500
loads: [rubric-authoring, traffic-light-decisions]
---

# Goal Harness

**TL;DR:** A long-running agent stops when an external reviewer says the rubric passes or a declared stop condition fires — never when it feels done — so every goal states five elements and a rubric before the loop starts.

## When to use

- The task is long, iterative, or unattended (an optimization initiative, a backfill, a design polish loop).
- "Make it better" is all you have — the harness forces a real definition of done first.
- An earlier run quit early or declared victory after two small edits.

## Steps

1. **Write the five elements** (template: [`assets/goal.template.md`](assets/goal.template.md)):

   ```yaml
   outcome:          # concrete target STATE — a number, threshold, or observable; never "better"
   verification:     # how we PROVE it — a tool run, a test, a rubric pass; never self-report
   constraints:      # what must NOT change — the scope fence
   iteration_policy: # what to log between attempts — {what changed, measured result, next-best idea}
   error_handling:   # when to PAUSE and REPORT (tool won't run, ideas exhausted, needs a human)
   rubric:           # path to the .rubric.md the reviewer scores against (rubric-authoring)
   ```

2. **Refuse a boundary-less goal.** If the outcome is an adjective or verification is missing, ask for the missing piece before looping.
3. **Run the loop**: the executor produces output; the reviewer scores every rubric row and runs the verification each round. Not met → name the specific gap and iterate. Met → stop.
4. **Use a tri-state verdict**: `PASS` (every MUST row green) · `ITERATE` (work remains the agent can do) · `BLOCKED` (only externally gated rows remain red). BLOCKED stops the loop.
5. **Pause and report** per `error_handling`: a stuck run is a 🔴 with a named blocker and unblock owner — never a silent give-up or a fake "done".
6. **Calibrate** on a new rubric: for the first 3–4 rounds a human compares the reviewer's verdict with the artifact; a mismatch means the rubric needs revising (`rubric-authoring`).
7. **Lint** goal files: `python3 06_Engine/harness/lint-skill.py goal-harness`.

## Output format

Initiative goals live at `00_Agentic-OS/initiatives/<slug>/<slug>.goal.md` with a sibling `<slug>.rubric.md`; project goals live beside the work in a hub. A worked contrast:

- ❌ "Make this project better." → two tiny edits and a victory lap.
- ✅ "Checkout responds in ≤ 0.2 s, measured by the speed-test tool over 5 runs; change only the checkout block and its tests; log change / measurement / next idea each try; if the tool won't run or ideas run out, stop and report what was tried, where it's stuck, and what you need."

## Watch-outs

- **Self-report is not verification.** "I improved it" must become a tool, test, or rubric pass.
- **No rubric on subjective work** lets the model grade its own homework.
- **The reviewer must score the whole rubric** — or print `NOT-SCORED (reason)` for each skipped row.
- **Iteration is not monotonic.** Round 10 can beat round 15; keep the best artifact, not the latest.
- **Looping forever.** Without `error_handling`, a blocked goal burns budget.

## See also

- Reference: [`references/REFERENCE.md`](references/REFERENCE.md) — why agents quit early, field-tested pitfalls, sources
- `rubric-authoring` (the rubric this loop scores against) · `traffic-light-decisions` (a stuck report is a 🔴)
- Agent: [`goal-runner.agent.md`](../../../03_Agents-and-Skills/Agents/goal-runner.agent.md) · Prompt: [`/optimize-loop`](../../../02_Workflows/prompts/optimize-loop.prompt.md)
- Framework: [`00_README/_frameworks/long-horizon-execution.md`](../../../00_README/_frameworks/long-horizon-execution.md)
Pliki w paczce (3)
  • SKILL.md 7 KB
  • assets/goal.template.md 1 KB
  • references/REFERENCE.md 3 KB

Wersje

W bazie nie ma innych skilli o nazwie goal-harness.

Migawka bazy z 13.11.2026 · wersja paczki 14656f68136f (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

  • 100goal-loop
    Pomaga pisać prompty /goal oraz wyjaśnia pętle agentów, w tym Ralph loop, razem z konfiguracją, monitorowaniem i rozwiązywaniem problemów.
  • 100goal-test
    Zamienia niejasne zadanie w testowalną definicję ukończenia i generuje wykonywalny skrypt goal-test, opcjonalnie z ograniczoną pętlą ponowień wokół agenta headless.
  • 100goal
    Porządkuje zadanie w sekcje TASK, WHY, OUTCOME, CONSTRAINTS i VERIFICATION, zanim zacznie się wykonanie, żeby praca miała jasną metę.
  • 90harness-loop
    Prowadzi duże zadanie jako pętlę z jednym sprawdzeniem, plikiem postępu i limitem tur, aż cel zostanie ukończony.
  • 100loop-controller
    Dodaje do zadania weryfikowalny warunek zatrzymania i obowiązkowy zestaw zabezpieczeń, by autonomiczna pętla zbiegała się zamiast się szarpać, i wybiera właściwy prymityw (/goal, /loop, hook Stop, pętla Ralph).
  • 100rubric-authoring
    Zamienia intuicyjne oceny jakości w punktowalną rubrykę w sześciu krokach, aby agent recenzent mógł oceniać subiektywną pracę według Twojego poziomu.

Uwagi i zgłoszenia

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