There's 3 ways to build a great Claude / Codex skill:
Define the process from scratch when you already understand the task, the inputs, and the desired output.
Extract the process from successful work after you have corrected the mistakes and accepted the result.
On Mac, use Codex Record & Replay when the workflow depends on repeatable actions across a browser, spreadsheet, or internal system.
I personally prefer running a chat session first to build a great output first, and then wrapping this together into an elegant skill.
(process #2 on the list above)
This works great because all the scripts and tool calls as part of the build gets referenced as context when you wrap it together into a skill!
Tip 2: Build the Workflow First
Before writing SKILL.md, run the workflow on a real company and review the result.
Give Codex the goal, the available data, the intended audience, and an example of strong output.
Open Codex in Plan mode and ask it to surface ambiguities through clarifying questions.
The sequence I use is:
Plan → Clarify → Approve → Execute → Review → Save
Once the result meets your standard, package the tested method and the corrections you made along the way.
Before Codex creates the files, answer a few practical questions:
When should the skill activate, and who will use the output?
What must the user provide, and can the skill access the required data legally and reliably?
Which sources take priority when two sources disagree?
Which steps require judgment, and which calculations should always run the same way?
What should the finished work look like, and what accepted example defines good?
What must be verified before delivery, and which decisions stay with the analyst?
These answers define the skill’s contract and give you something concrete to test.
Confirm the source workflow before building the rest because polished instructions cannot fix unavailable or unreliable data.
Tip 3: My Starting Prompt
If you want to build a skill from scratch rather than wrap an existing output into a skill, here's a simplified starter prompt I used to build my earnings skill:
I want to build a reusable Codex skill for public-equity investors who analyze earnings calls.
The user will provide a company name or ticker.
Analyze the latest completed fiscal earnings call and the three completed calls before it.
Prioritize company investor-relations materials and official filings.
Use only transcripts that can be accessed legally, record the source of each call, and leave missing calls visible rather than filling the gaps.
Produce a self-contained HTML report covering what changed, KPI and guidance history, management-language trends, Q&A themes, risks, clickable sources, and limitations.
Use scripts for calculations that must be repeatable.
Keep buy or sell recommendations, price targets, and position sizing outside the skill.
Before creating any files, ask me clarifying questions about source conflicts, edge cases, the output standard, and the review process.
Then give me a plan for approval.
Tip 4: Make SKILL.md the router
For a complex workflow, the main SKILL.md file should act as a router.
It explains when the skill should activate, what the workflow does, and which supporting files Codex should load.
Then the nuances (edge cases, instructions, any python scripts, etc) get routed to other subfolders.
The detailed logic belongs in the supporting folders.
I like to have this as the .zip folder architecture for finance skills:
(Agents btw is Codex specific to call for subagent architecture, you need to swap that out if you use Claude Code)
What my Earnings Skill Does:
Earnings analysis looks simple until you try to make it repeatable.
Transcripts can be missing, KPI definitions can change, and guidance may appear in the prepared remarks, the Q&A, or a separate filing.
A generic summary can sound polished while missing the information an investor needs.
I built the Earnings Call Analyzer to handle those problems explicitly.
The skill collects the four most recent completed fiscal earnings calls, records where each transcript came from, and keeps missing calls visible as gaps.
It cross-checks important numerical claims against official company disclosures.
I also added in sentiment analysis - something that manual scrubbing of earnings calls can't handle.
The finished HTML report includes:
An executive summary of what changed.
KPI and guidance history.
Management-language and hedge-word trends.
Q&A pressure points and recurring themes.
Risks and items worth monitoring.
Clickable citations, methods, and limitations.
Some parts require investor judgment, while calculations need consistent treatment on every run.
For example, a script counts and normalizes hedge language before Codex interprets the pattern alongside the transcript and official disclosures.
The headline language metrics also focus on named executives rather than mixing management commentary with analyst questions or operator text.
Download my Earnings Call Skill:
I hosted my earnings call skill on GitHub for you to install.
