agent-audit
/agent-audit
Sprawdza wiersz po wierszu, co agent GTM przygotował, i decyduje przy każdym kroku, czy go zostawić, zastąpić regułą albo postawić przy nim człowieka.
- Autor
- derrickikenga ↗
- Repozytorium
- gtm-skills ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 11.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 12
- Skrypty
- 2
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 11.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 2 skrypty w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: scripts/audit.mjs, scripts/audit.test.mjs
Co robi i kiedy
Sprawdza wiersz po wierszu, co agent GTM przygotował, i decyduje przy każdym kroku, czy go zostawić, zastąpić regułą albo postawić przy nim człowieka.
Oryginalny opis (EN)
Audit what a GTM agent produced, row by row. Find the rows it got wrong, then decide for each step whether to keep the agent, replace it with a rule or formula, or put a person in front of it, plus a plan to monitor it. Use when someone asks if their AI column, Claygent, n8n agent or AI SDR is accurate, says an agent qualifies or scores leads badly, asks whether to use AI or a formula, wants human-in-the-loop for outbound, or wants to check AI-written emails before they send. Works from a CSV export; it does not log in to any tool.
Jak działa
- Workflow
- Rules
Podgląd SKILL.md
--- name: agent-audit description: Audit what a GTM agent produced, row by row. Find the rows it got wrong, then decide for each step whether to keep the agent, replace it with a rule or formula, or put a person in front of it, plus a plan to monitor it. Use when someone asks if their AI column, Claygent, n8n agent or AI SDR is accurate, says an agent qualifies or scores leads badly, asks whether to use AI or a formula, wants human-in-the-loop for outbound, or wants to check AI-written emails before they send. Works from a CSV export; it does not log in to any tool. --- # Agent audit An agent ran for a while: it enriched, classified, scored or wrote emails. This skill finds the rows it got wrong and decides what each step needs: - **keep the agent**, and monitor it - **replace it with a rule**, because a formula gives the same answer for free, every time - **add a human review gate**, because the step sends or writes outside your systems and makes mistakes Code finds the suspect rows. A person decides what is right. The script counts. ## Workflow 1. **Get the rows and the steps.** Read `references/intake.md`. You need: - a CSV of what the agent produced, with its inputs - for each AI step: the prompt, the column it wrote, and whether it sends or writes outside (email, CRM) 2. **Write `steps.json`.** Format: `references/steps-format.md`. Start from `examples/clay-hr-fit/steps.json` or `examples/outbound-drafts/steps.json`. - For each classify step, write one candidate rule: the formula a person would write from the instruction, with only the columns in the table. - ⛔ Write the rule from the definition, never from the agent's answers. A rule tuned to the agent copies its mistakes. 3. **Run the check.** ```bash node scripts/audit.mjs check rows.csv steps.json --out review.csv ``` The review sheet holds every flagged row, plus a random sample of clean rows (20 by default, `--sample`). The sample shows the errors the checks cannot see. 4. **Review.** A person sets `verdict` to `right`, `wrong` or `unsure` on each row. When a row is wrong, they put the right answer in `correct_value`. - If the user asks you to review, read each row's input and decide. Write "reviewed by Claude" in `note`. - Then say in the report that a person did not review it. - ⛔ Never fill a verdict you cannot support from the row's own data. 5. **Score and report.** ```bash node scripts/audit.mjs score review.csv steps.json --rows rows.csv ``` - Lead with the table: each step, its error rate and its decision. - Then, for each step: why, the rows that were wrong and what they had in common, and the plan to monitor it. - Read `references/decisions.md` before you add advice the script does not give. ## Rules - **Every error rate comes from reviewed rows.** Under 30 reviewed rows, say "direction, not measurement". - **A flag is not an error.** A rule can be wrong too. Only the verdict column counts. - **Never send, post or write anything during the audit.** Drafts stay drafts. - **Keep personal data local.** Rows can hold names and emails. Do not paste them into other tools or reports. - **Say when the agent did well.** A short report that says "keep it" is a good result.
Pliki w paczce (12)
- SKILL.md 3 KB
- examples/clay-hr-fit/review-filled.csv 5 KB
- examples/clay-hr-fit/rows.csv 1 KB
- examples/clay-hr-fit/steps.json 1 KB
- examples/outbound-drafts/review-filled.csv 2 KB
- examples/outbound-drafts/rows.csv 2 KB
- examples/outbound-drafts/steps.json 1 KB
- references/decisions.md 2 KB
- references/intake.md 1 KB
- references/steps-format.md 2 KB
- scripts/audit.mjs 20 KB
- scripts/audit.test.mjs 5 KB
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · derrickikenga/gtm-skills | 100 | MIT |
| gmen1057/agent-audit | 100 | MIT |
| bdfinst/agentic-dev-team | 95 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 11042bf02005 (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
- 90leads-with-emails
Buduje listę leadów B2B, w której każdy wiersz ma działający adres e-mail, także z zapisanego wyszukiwania Apollo. - 100ai-prompts-toolkit
Dostarcza bibliotekę promptów AI dla GTM i wzorce pętli promptów: badania Claygent, punktacja LLM, szkice cold maili, klasyfikacja odpowiedzi, briefy kont i iteracja do jakości w Clay, n8n lub Cursor. - 100rows
Wykonuje jedno zadanie w każdym wierszu arkusza: kategoryzuje, wyciąga dane z tekstu, ocenia lub tłumaczy i wpisuje wyniki w nową kolumnę. - 100sales-os
Obsługuje cały proces outbound: wyszukuje prospektów z intencją zakupu, sprawdza ich wobec ICP, pisze wiadomości na LinkedIn i raportuje konwersje.