Skills and Project Records, a Prepared Example

Prepared fictional example. Northstar, Alex, Priya, Sam, Jordan, Maya and Meridian Manufacturing are invented for this course. These records, prompts and plans are authored teaching material, not output from an observed agent run. Nothing has been sent, saved or scheduled.

This is the text of the Section 2 practice page. The same page, with its four tabs, is in the starter zip as practice/skills-and-project-records.html.

Three scopes keep background organized: Company, then Department (Delivery Operations), then Project (Northstar Onboarding). Explicit retrieval and update confirm what the agent is reading.

Three records in Northstar Onboarding

1. Project instructions. Check delivery dependencies before drafting client updates; keep confirmed plans separate from items awaiting confirmation. Reusable follow-up methods stay in skills.
2. Current facts (updated October 5). Jordan confirmed 18 of 22 client staff roles for Tuesday’s setup; the remaining 4 specialist roles wait until Thursday’s working session. Team: Priya leads delivery; Sam coordinates onboarding.
3. Decision record (October 5). Start Tuesday’s setup for the 18 confirmed staff members and hold the 4 specialist accounts until Thursday so onboarding proceeds without guessing security permissions.

Explicit retrieval and update check

• Retrieval confirmation. Agent reports reading current-facts.md and decision-record-2026-10-05.md before starting the next task.
• Before update. “Jordan intends to send the role list by 3 p.m. Friday, subject to operations-manager confirmation.”
• After update. “18 of 22 client roles confirmed for Tuesday Oct 6 setup; 4 specialist roles held until Thursday Oct 8 pending operations-manager sign-off.”

Paste the complete capture request into your successful conversation so the agent generates an adaptable prompt with zero blank placeholders, then test it in Work and capture it with @skill-creator.

Act as an expert prompt engineer. Examine this successful conversation, including the original goal, the approach we used, my corrections, the useful final output and the criteria that made the result useful. Use the relevant role, business context, preferences and source material I have supplied here.

Create a comprehensive, detailed, thorough, holistic and versatile prompt that preserves this successful method and can guide another AI agent through similar work. Cover the relevant purpose, specialist perspective, usable expert methods, reasoning, inputs, output criteria, important exceptions and action boundaries. Be thorough about what matters to this job, without unrelated padding or a requirement that every prompt be long.

Separate stable instructions from changing task facts. Explain how the agent should use newly supplied material and adapt when the audience, circumstances or available information changes. Do not embed this task's temporary facts as permanent assumptions. Preserve the reason behind important choices and corrections so the method can remain useful in a different situation.

Infer information only when the conversation and supplied context support it. If an essential missing fact would materially change the method, ask me naturally before finalizing. Do not create placeholders, blank form fields or a template that makes me repeatedly rewrite my context. Ask for fresh task material when the work needs it, and retrieve relevant records only when access is available and permitted.

For substantial or unfamiliar work, have the executing agent propose the approach and ask consequential questions before execution, so I can refine the plan. For familiar small work, let the agent use an accepted method directly. Include checks that judge the result against the actual goal. Preserve appropriate expertise, privacy, permissions, decision ownership and limits on external actions; do not invent tool access or authority.

Return the complete ready-to-use prompt. Briefly explain any important choices in how you generalized the method and flag unresolved limits. Do not claim that the prompt has already been tested, that the next task will succeed or that a skill has already been created.

Testing @meeting-follow-up on Wednesday’s technical check-in exposes one recurring preference: naming a single confirmation owner when colleagues share a dependency.

When an action or preparation dependency involves more than one colleague, state who is assisting and name the single owner responsible for sending confirmation back to the leader.

Thursday’s security check and working session run at 2 p.m. on October 8; Sam will attach session links by 1 p.m. and confirm completion back to Alex by 1:30 p.m.

Maya corrects an unfamiliar RFP evaluation plan before execution.

Maya’s pre-execution correction on Meridian Manufacturing

• Flawed initial plan: Treated total engineering headcount (6 engineers) as interchangeable with senior integration specialist availability.
• Corrected plan before execution: Check the 6-week migration timeline specifically against senior integration specialist availability in weeks 3 and 4, and include a phased-start option moving custom integration to week 5.

Meetings and follow-through coordinates focused skills with explicit handoffs.