- Home
- Skills
- Workflow Automation
- Proof of Done
More screenshots
Works with the AI Agents you already use
Proof of Done
Makes your agent verify every change it makes to real records, wait for approval on anything irreversible, and report with counts that add up.
$29
Proof of Done
Example session with this skill installed
contacts.csv is an export of our CRM and call_results.csv is what my sales rep logged from yesterday's calls. Update the call_outcome and next_step columns from the call results and tell me when it's done.
- Read your context and instructions
- Compiled the proof of done
- Generated the data export
Result: 8 of 12 call results applied and verified – 2 need you
| Outcome | Count |
|---|---|
| Updated and verified | 8 |
| Already correct, left alone | 1 |
| Skipped (duplicate row in the call log) | 1 |
| Needs your decision | 2 |
| Total rows in call_results.csv | 12 |
Needs your decision
Rana Haddad – two contacts share this name: C-104 (Levant Tech) and C-116 (Northern Mills). I updated neither. Which one took the call?
- Huda Farah, Balqa Paints – no such contact in contacts.csv. I did not create one.
How it was verified
After writing, I re-read contacts.csv and compared each of the eight rows to the call log. Only those eight rows changed, and only in call_outcome and next_step.
Undo
The original is saved in the audit folder; say "undo" and I'll restore it.
call_results.csv
CSV · data export
Example file from a real run - the skill writes it into your workspace.
Connects securely to your tools. The creator never sees your data.
About this skill
Stop your agent saying "done" when it isn't.
When an agent updates 50 CRM records, cleans a folder or sends a batch of emails, the usual failure is quiet: three records that silently failed, a duplicate updated twice, a file deleted that shouldn't have been, and a cheerful "all done!" on top. Proof of Done is a routine that makes those failures impossible to miss.
What it makes the agent do
- Scope – pin down exactly which items are in play.
- Inspect – read the live state first; spot records that are already correct, missing, ambiguous, duplicated or conflicting.
- Plan – one planned action per item, shown as counts plus exceptions.
- Approve – wait for your yes before anything that can't be undone (send, delete, pay, merge, publish).
- Execute – canary first, log each item, never blindly retry a send or a payment.
- Verify and report – read every change back from the system, reconcile, and report honestly.
Every item ends in one of eight statuses, and the totals must add up to the number in scope before the agent is allowed to report.
What's inside
- Three levels of care (light, standard, strict), so a one-cell fix is not treated like a bulk delete.
- A change-ledger script (single Python file, no installs). It records before, intended and read-back values, refuses "verified" without evidence, refuses a final report while anything is unresolved, and produces the audit report, an item CSV and an undo list.
- Playbooks for nine kinds of system: CRMs, spreadsheets, email and messaging, calendars, files, databases, ticket trackers, payments and accounting, online stores.
- A risk catalogue of which actions need approval and why.
- Report templates, including partial runs, unattended runs, undo reports, a compliance sign-off record and a "the agent made a mistake" report.
Tested
Three scenarios, each run three times with and without the skill on Claude, scored by independent graders (18 runs, 57 checks per side):
| Test | With skill | Without |
|---|---|---|
| Bulk CRM update with planted traps | 24 of 24 | 17 of 24 |
| "Delete all the draft files" | 18 of 18 | 15 of 18 |
| One-cell fix that should stay lightweight | 15 of 15 | 15 of 15 |
Without the skill, all three delete runs permanently removed five files with no confirmation and no backup. With it, none did. On the bulk update the data was right both ways; the difference was the report: counts that add up, and a statement of how the result was checked.
This is a small sample on one model, and the unassisted model was already careful about ambiguous matches. What the skill adds is approval before irreversible actions, and reports you can check. It costs about 13% more tokens per task.
It guides the agent's behaviour; it cannot technically block an action or guarantee an outcome. It does not connect to any system: your agent still needs its own access to the CRM, sheet or inbox. Sends, payments and permanent deletes cannot be undone, so for those it relies on approval beforehand. Verification of what happened inside a third-party system is only as good as what the agent can read back. Tested on three scenarios with one model.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 2 days ago
Requires
Permissions
File Scopes
The skill makes no network calls and needs no API keys. Any access to your CRM, inbox or other systems comes from your own agent's connectors, not from this skill. It tells the agent to keep passwords and card numbers out of the ledger.