What should the process do when something fails?
Before scheduling, test these cases using a copy of the workshop data. A useful automation can also stop and explain a problem.
- Empty file
- Stop the report and flag missing data. Do not send a zero-sales report.
- Duplicate document
- Count an identical record once. Flag conflicting versions of the same ID for clarification.
- Source unavailable
- Notify the owner that the run failed. Do not silently substitute an old file.
- Retry after failure
- Check whether the previous attempt already created or sent a result. Use a period identifier to avoid sending twice.
Record the owner, time zone, runtime, failure notification and stop procedure. A skill describes the procedure; a schedule specifies when to run it.
Module D
Reusability and the cloud - from a good prompt to a process that runs itself
A good prompt nobody saved dies with the conversation. This module is about turning it into a Skill, moving skills between tools, running them in the cloud on a schedule, and branching conversations so an experiment does not cost you a good thread.
A Skill is a recipe card, not a prompt
A skill is a reusable instruction: company definitions, steps, output format and quality checks (QA). The team shares the same rules, but results still depend on data, tools and permissions.
You have done it three times
Once is an experiment, twice a coincidence, three times is a process.
You keep making the same corrections
Every run you re-add the same notes about format or tone.
Someone asks "how did you do that"
A Skill is the answer you do not have to repeat on Slack.
The result must look the same
A report with the same structure each month can be compared. A different one each time cannot.
What a finished Skill looks like
Dla zaawansowanych: przykładowy prompt +
name: weekly-sales-report when to use: whenever I ask for a weekly sales summary our definitions: - "sales" = invoices issued, not orders - "new customer" = first invoice within the last 90 days - the week runs Monday-Sunday steps: 1. pull data from the [SOURCE] sheet for last week and the week before 2. compute: revenue, deal count, average value, new customers 3. compare with the previous week and with the monthly target 4. list at most 3 decisions to make (what, who, by when) output format: - a 5-sentence summary - a plan vs. actual table - a "decisions" section - an "assumptions and data gaps" section QA: - the parts must sum to the whole - the same date range as last week - every number has a source do not: - do not forward the report, hand it to me for approval - do not fill missing days with an estimate
Moving a skill from Claude into ChatGPT Work
Use the Claude instruction as a starting point. After moving it, review tools, dependencies, files and permissions, then compare outputs on known data. Uploading the file alone does not establish compatibility.
1. Download the skill file
In Claude (Chat or Co-work) skills are plain text files. You download the skill folder with its instructions.
2. Drop it into ChatGPT Work
Upload the file to a conversation and ask for it to be converted into a Skill that runs in Work.
3. Check the tool names
Check tool names, permissions, reference files and required scripts. Flag anything the new environment does not support.
4. Run it on old data
Run the skill on an example you already know and compare the result with what you used to get.
5. Add your own rules
Your metric definitions and hard limits get written in once, at import - not on every run.
Dla zaawansowanych: przykładowy prompt +
I uploaded a skill file from another AI tool. TASK: convert it into a Skill that runs here. DO THIS: 1. summarize in 5 bullets what the skill does - I will verify it, 2. list every reference to a tool or permission and propose an equivalent available in my configuration, 3. flag the steps that cannot be reproduced 1:1 and say why, 4. hand back the finished skill content, 5. propose one test case I can check it with. Do not change definitions or numeric thresholds - if something looks wrong, report it separately instead of silently fixing it.
Before importing someone else's skill: read all of it. A skill is an instruction the agent will follow - treat a foreign file the way you treat a foreign script.
Skill + schedule = a process without you
Check the execution environment first. If the task runs in the cloud and all its sources are available there, your laptop can be off. A task using local files or apps may need the computer to stay on. Test the schedule, time zone and how to stop the process.
Hourly
Monitoring: a mailbox, a form, a ticket queue. Decide what happens when nothing changed - ideally nothing.
Daily
The morning review, an end-of-day summary, a check on yesterday's data.
Weekly
Reports and reviews - the only rhythm where you see a trend rather than noise.
Monthly
Closings, settlements, and a review of what is still running at all.
Dla zaawansowanych: przykładowy prompt +
Every Monday at 8:00 run the "weekly-sales-report" skill on the data in [DRIVE FOLDER] for the previous week. OUTPUT: a document saved in the same folder, named "report-YYYY-MM-DD", plus a message to me with a three-sentence summary and a link to the file. RULES: - if last week's data is missing, tell me and do NOT build a report from an incomplete period, - do not send the report to anyone but me, - show me the first four runs for approval before saving.
Availability depends on your plan, region, workspace and administrator configuration.
Branching a conversation - risk-free experiments
A long conversation with well-built context is valuable and easy to ruin with one unfortunate instruction. Branching lets you fork from a chosen message and try another idea without touching the original.
Go back to a message
Pick the point in the conversation where things were still right and start a new branch from there.
The original stays
A branch does not erase what came before. You can compare two versions instead of guessing which was better.
Test variants
Same context, three different offer tones or three report structures - without pasting the context again.
Rescue derailed threads
When a conversation went wrong and each fix makes it worse - go back to the fork instead of starting over.
In practice: when a brief is ready and you want to compare variants, fork three branches from the same point and judge the results side by side. Save the winner as a Skill - that is how an experiment becomes a standard.
Codex C4 - Skills
Skills - your company's know-how in Codex
In short ↳A skill contains task instructions and may include reference files or scripts. Names such as /remotion-captions appear in this screenshot; invocation depends on the app version. Read the instructions and inspect the result.
Where to find it: Codex → Plugins → Skills
Select the screenshot to enlarge it.
What you see on screen
- 1
Skills tab
The list of every procedure Codex knows. Each has a name starting with a slash and one line saying what it does.
- 2
Search skills
With a few dozen skills, searching is faster than scrolling. Name them by task, not by tool.
- 3
Installed
Installed skills and their status marks. A checkmark does not prove that a skill ran or that the result is correct.
- 4
Personal
Your own skills - your way of writing offers, your report format, your checklist.
- 5
System
The System category shown here. Inspect the source and instructions; this label alone does not imply approval by your company.
- 6
Add
Adds a new skill. In practice: a short text file - purpose, steps, rules, what the result must look like.
- 7
Remotion examples
The screenshot lists instructions for creating videos, captions and maps. Installation alone does not ensure all rendering dependencies or integrations are available.
How to use it - in plain words
How to write a good skill: one job, five to ten numbered steps, a short list of rules ("never send anything yourself", "always ask for the client name") and an example of the finished result.
A skill is worth writing the second time you explain the same thing to Codex. From then on you type one slash instead of a paragraph.
For advanced users: example prompt+
Skill: /oferta-standard Purpose: prepare a client offer in our format. Steps: 1. Ask for: client, scope, deadline, budget. 2. Pull prices from the price list in the project. 3. Write the offer: problem, solution, scope, price, deadline, next step. Rules: no promises about dates we have not confirmed. Never send it - leave a draft.
Try it on your account
Find a skill by name and read its instructions. A list checkmark is not proof that the procedure ran in a conversation.
Codex C5 - Scheduled tasks
Scheduled - work that runs without you
In short ↳You set a task once and Codex repeats it: a Monday audit, a daily report, monitoring a page for changes. The result waits for you in a chat.
Where to find it: Codex → Scheduled
Select the screenshot to enlarge it.
What you see on screen
- 1
Create
A new recurring task: what to do, how often, where to leave the result.
- 2
Monitoring
The first entry shows Monitoring and a next-run time. Whether unchanged results generate notifications depends on the task configuration.
- 3
Weekly task
'Mondays at 8:00' - the classic: an audit, a sales summary, a list of things left hanging.
- 4
Daily task
A short daily briefing at a fixed hour. Better than a long weekly report nobody reads.
- 5
Active / Paused / Completed
Pause tasks during holidays and delete ones you stopped reading - otherwise they only burn your limit.
- 6
Next run
Shows the planned next execution. It does not confirm that the previous run succeeded; open the task and inspect its result or error.
- 7
All and search
All lists tasks across states. Search filters names; Active, Paused and Completed narrow the list by state.
- 8
Cloud symbol and schedules
The screenshot contains cloud marks and relative run times. Inspect task details for execution requirements, time zone and notifications; the symbol alone does not guarantee a run while your computer is off.
How to use it - in plain words
Start with one task that you currently do by hand every week. Run it for a month, then decide whether it is worth keeping.
Every scheduled task needs an owner - a person who reads the result. A report nobody opens is not automation, it is noise.
For advanced users: example prompt+
Scheduled task: Mondays at 8:00. 1. Check what changed in the project over the last 7 days. 2. List: what was finished, what is stuck, what is at risk. 3. Maximum one page, no jargon, ready to paste into Slack.
Try it on your account
Find Paused, Next run and Create. Check your task time and time zone before saving it.