forge-agent
/forge-agent
Przygotowuje brief, buduje i sprawdza nowego agenta AI, zanim trafi do użycia, z ewaluacją i pytaniami o cel.
- Autor
- mjha-arya ↗
- Repozytorium
- claude-code-skills ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 10.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 1
Brak sygnałów. Repozytorium zmienione 10.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Przygotowuje brief, buduje i sprawdza nowego agenta AI, zanim trafi do użycia, z ewaluacją i pytaniami o cel.
Wyzwalacze z opisu
Argumenty
/forge-agent One-line idea of the agent, however rough
Oryginalny opis (EN)
Brief, build, and prove a new agent (Claude Code subagent/skill, Cursor agent, Agent SDK, Python/LangGraph, scheduled job) before it ships. Use when the user types /forge-agent, says "create/build/design/spec/scaffold an agent", "brief an agent", "ask me the agent questions", or gives a one-liner naming a stack ("SDK + SQLite to do X") but no outcome or eval.
Jak działa
- Gate 1: Brief (no build without it)
- Gate 2: Build
- Gate 3: Prove
- Close
Podgląd SKILL.md
---
name: forge-agent
description: Brief, build, and prove a new agent (Claude Code subagent/skill, Cursor agent, Agent SDK, Python/LangGraph, scheduled job) before it ships. Use when the user types /forge-agent, says "create/build/design/spec/scaffold an agent", "brief an agent", "ask me the agent questions", or gives a one-liner naming a stack ("SDK + SQLite to do X") but no outcome or eval.
argument-hint: "One-line idea of the agent, however rough"
---
Works well by voice: ask **at most 4 numbered questions per round**, each with a **default**, so
replies like "default" or "1 default, 2 is …" work. One round per message, then wait. Pre-fill whatever
the ask already answers and mark it "taken". If a name or number looks odd (dictation garbles them),
ask ONE question and never guess. Outcome before stack: park any named tech ("Noted SQLite, deciding
at step 3"). Choose the model only after the golden cases exist.
**Config:** `AGENTS_ROOT` is where your agents live (default: `./agents`). `CONSTITUTION` is your rules
file (default: [Constitution of Agents](https://github.com/mjha-arya/constitution-of-agents) `AGENTS.md`)
and `GLOSSARY` is its agent registry.
## Gate 1: Brief (no build without it)
1. **Reuse:** read `GLOSSARY`. If an agent already covers the job, propose extending it instead.
Report in one line how many agents are still `Authored, not verified` (grep `**Status:**` in
`AGENTS_ROOT/**/README.md`).
2. **Job** (outcome, not stack) · **Done when** (a number; *default: at least 4/5 golden pass, zero
safety violations*) · **5 golden cases**, each input → ideal output, at least 1 edge case and at least
1 adversarial (*default: you draft, the user corrects*).
3. **Data** in → out, with exact paths · what goes stale (*default: re-fetch each run, cache at most 24 h*) ·
**persist?** If nothing needs to persist, no DB (*default: stdlib `sqlite3`*) · memory write-back.
Procedural changes go to `PROPOSALS.md` for the user's yes.
4. **Tools:** read set, write = exact paths only, and outward actions (post/email/schedule/deploy) need
the user's yes · PII/secret check on output · who owns it when it breaks (*default: the user, via
their notification channel*).
5. **Runtime + REAL auth:** `claude -p` | subagent/skill | Agent SDK | LangGraph. **Test the auth now.**
Placeholder keys are the most common reason an agent never runs. Single agent or a pattern ·
`max_turns` · $/run (*default: `claude -p` or subagent, single, 30 turns, $1*).
6. **Failure → behaviour:** login wall, empty data, rate limit, tool error, low confidence
(*default: retry at most 3 times, exit 2, send a notification, JSONL trace in `data/runs/`*).
→ Write `AGENTS_ROOT/<Name>/BRIEF.md` + `evals/golden.jsonl`. Read it back in 10 lines or fewer and get
the user's yes. No golden cases → mark it `DRAFT: no eval` and stop.
## Gate 2: Build
7. Read `CONSTITUTION` in full and pick the runtime by its rules (one line).
8. The soul **embeds** the applicable rules (standalone, no links): purpose, inputs, hard rules, a
`needs_human` list, and an exact return format ending in NOT CHECKED.
9. Build it yourself, or hand large builds to another coding agent with the brief path + spec, then
review the diff.
## Gate 3: Prove
10. Run **every** golden case on the real model/runtime. Log each to `data/runs/*.jsonl`.
11. Score: deterministic checks + tool-call count (no loops, no duplicate calls) + an LLM judge only
for the fuzzy part.
12. Fail → fix the soul → commit `"failure fixed → expected change"` → rerun. At most 3 loops, then
report the blocker honestly.
13. Pass bar met → README `**Status:** Verified (<date>, <score>)`. Otherwise `Authored, not verified`.
Never claim a run that didn't happen.
## Close
14. Add or update the agent's row in `GLOSSARY` and commit it.
15. Reply in one line: path · score · NOT CHECKED.
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · mjha-arya/claude-code-skills | 100 | MIT |
| aaddrick/ticketmill | 100 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 22af0e7a3830 (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
- 100forge-install
Przeprowadza pierwszą instalację wtyczki Forge Agent Lens dla Claude Code, w tym zależności i konfigurację projektu, by śledzić sesje. - 100create-agent
Tworzy nowego agenta-współpracownika, czyli zaplanowaną personę Claude uruchamianą harmonogramem, na polecenie typu „stwórz agenta”. - 95skill-forge
Tworzy i sprawdza nowe skille Rune metodą testów przed zapisem, zanim skill trafi do użycia. - 100tool-forge
Wykrywa lukę w narzędziach (TOOL_GAP), tworzy narzędzia i szkielety skilli dla bieżącego agenta i rejestruje je w katalogu możliwości. - 100forge-config
Odczytuje i zmienia ustawienia wtyczki Forge Agent Lens dla Claude Code, takie jak projekt i klucz wandb, oraz restartuje demona, by zastosować zmiany. - 100agent-forge
Buduje, konfiguruje i naprawia agentów zadaniowych i aplikacje agentowe, w tym trwałe, wywołujące narzędzia i zaplanowane, albo robi przegląd wskazanej zmiany.