Use it when
- The work happens frequently and follows recognizable criteria.
- The team knows the inputs, exceptions and output of a good result.
- The method should work across projects or conversations.
Do not start here when
- Every run is fundamentally different and depends on tacit context.
- A person has not tested the routine yet.
- The plan is a pile of generic instructions with no examples, limits or quality criteria.
A skill is not the same as automation
A skill packages instructions and resources that teach AI a method: what to consult, which order to follow, when to stop and what output to produce. A connector or MCP provides data and actions. They may work together, but solve different problems.
If the team repeatedly explains how to review a proposal, you may need a skill. If AI cannot read the CRM, you have an access problem.
1. Meeting preparation
Inputs include the event, attendees, prior decisions and related documents. The method finds unresolved work, conflicts and decisions to make. The output is a short brief with objective, risks and a proposed agenda.
It should stop on an ambiguous event or person instead of guessing.
2. Commercial proposal review
Compare the proposal with approved scope, commercial policy and company criteria. Surface conflicts in deliverables, price, timing, dependencies and terms. Prepare questions, but never invent a condition or approve a discount.
3. Weekly pipeline review
Inspect opportunities, last interaction, stage, value and next step. Identify stalled deals, missing fields and overdue commitments. The operation must define what “stalled” means.
4. Client onboarding
Turn contract, plan, owners and requirements into a checklist with dependencies and missing access. Sensitive data should appear only for the person handling that step. Scope or deadline changes require accountable approval.
5. Delivery quality control
Compare the final artifact with the brief and acceptance criteria. Check requirements, sources, links, format and points that need human inspection. Objective checks outperform instructions such as “make it perfect.”
6. Incident triage
Organize the report, time, affected service, allowed logs and impact. Separate evidence from hypotheses, propose severity and identify the next tests. Production changes and customer communication remain subject to the incident policy.
7. Weekly management brief
Compare defined metrics, deliveries, risks and pending decisions with the prior week. Produce one page of facts, questions and owners. Every number should retain its source and period.
Test the method, not the wording
Give the same inputs to two people using the skill. Check whether they reach the same decisions, respect the same limits and produce usable output. Record every place they had to interpret or ask for help.
Refine with real examples and exceptions. Distribution comes after the workflow is reliable.
OpenAI's GTM Assistant incorporated the habits of top sales representatives
OpenAI describes an internal Slack assistant that prepares meetings and answers product questions using account history, call notes, Salesforce activity and release updates with traceable source links.
Representatives could spend an hour preparing for a thirty-minute call.
The assistant produced daily meeting briefs and recaps across several sources.
Top representatives helped define what a good brief and answer looked like.
Evals, corrections and improvements spread that operating standard across the team.
Connections supplied data. The differentiator was a shared definition of quality. A good skill preserves that sequence, rubric, limits and stop conditions.
Turn a routine into a skill without creating an unmaintainable manual
Choose one routine and work from a real execution. Start with reliable work, not with the skill file.
1. Collect one good and one bad example
Remove data that cannot circulate and mark the operational differences.
Evidence of the standard the skill should preserve.
2. Define the contract
State triggers, required inputs, output and stop-and-ask conditions.
A clear activation and completion boundary.
3. Convert expertise into decisions
Replace vague quality language with observable checks.
A sequence another person can review.
4. Separate instructions and references
Keep the procedure short and move policies, examples and schemas into supporting files.
Less context used on each run.
5. Test activation and refusal
Include complete, incomplete, out-of-scope and dangerous requests.
Positive and negative test cases.
6. Record recurring corrections
Turn repeated feedback into a criterion, example or test.
A skill that improves without becoming bloated.
Generate the first specification for your skill
The full pack will include a working file structure, reference templates, test cases and an installable starting point.
Help me turn a company routine into an AI skill specification. ROUTINE [name and frequency] EXAMPLE OF GOOD WORK [describe or paste one example] ALLOWED SOURCES [documents, systems and references] PROHIBITED ACTIONS [what AI must never do without approval] Create: 1. objective and activation triggers; 2. required inputs; 3. work sequence; 4. quality criteria; 5. stop-and-ask conditions; 6. exact output format; 7. three positive tests; 8. three negative tests; 9. facts that must never be invented. Replace generic instructions with observable decisions.
The guide stays open. You pay for the shortcut.
Join the early list. We will use demand to decide which pack should be released first and what it must include.
- Editable SKILL.md
- Reference, asset and script templates
- Quality rubric
- Activation, refusal and exception tests
- Version and maintenance checklist
Verify at the source
Product capabilities and policies change. These are the official references reviewed for this guide.