Yueseryueser.com

AI / Skills · multi-step workflows

Skills · multi-step workflows

Things a single prompt can’t do — these chain several steps together

1Turn a vague request into an actionable plan

For when you have picked up a new task and the other side has not thought it through either

  1. Have the AI ask you 5 questions first, to fill in the blanks in the request (prompt: I need to [task]. Ask me the 5 most important questions first, and do not propose a solution yet.)
  2. Write your answers up as a single background paragraph, then have it produce three plans at different scopes (minimum viable / normal / full).
  3. Pick one plan and have it break that into steps that can each be verified on their own, with a definition of done for each.
  4. Have it name the step most likely to stall, and the fallback for when it does.

2Debug a new error

So you do not burn half an hour heading in the wrong direction

  1. Paste only the error and the stack trace at first, and have it rank the possible causes — hold the code back so it does not get anchored by it.
  2. For the top-ranked cause, ask for a way to verify it with a single command.
  3. Once you have verified it, paste the minimal reproduction and have it pinpoint the line.
  4. After the fix, have it summarize: what class of bug this was, how to prevent it next time, and whether the same pattern exists elsewhere.

3Writing a Document That Will Be Reviewed

Poke holes in it yourself before you hand it over

  1. Have it list the questions this document must answer, from the target reader's point of view
  2. Write your draft following the order of those questions
  3. Have it list objections as the harshest possible reviewer (prompt: assume you're going to reject this document — what are your reasons?)
  4. Add a paragraph addressing each objection, then cut all the defensive filler

4Getting Up to Speed in an Unfamiliar Field

Build enough understanding to hold a conversation within two hours

  1. Have it give the field's five core questions and three common misconceptions
  2. For each core question, have it explain why the question matters and what the current mainstream answer is
  3. Have it list five key terms in the field with their precise meanings (be careful to distinguish near-synonyms)
  4. Finally, have it produce a 10-question self-test — whatever you can't answer is your gap

5Turn a Failure into Reusable Lessons

Post-mortems easily turn into blame-shifting or self-congratulation

  1. Start with a bare timeline of facts, no judgments, and have it point out what is missing from the timeline
  2. Have it list 'what information was available at the time that we did not use'
  3. Classify the improvement items by 'can this be changed next time' (can change / cannot change)
  4. Keep only the changeable part, and write each item as one concrete action + a checkpoint

6Set Up Recurring Tasks for AI (Build a Prompt Template)

When you repeat the same kind of task, turn the prompt into a reusable asset

  1. Do the same kind of task manually 3 times, and record how you described the requirements each time
  2. Extract the parts that repeat across all 3 into fixed sections (role / constraints / output format)
  3. Mark the parts that change each time as variables (using [square brackets] as placeholders)
  4. Test the template on a 4th new task, filling in blanks without changing the structure; if results are unstable, add one more constraint