Turn the way this person already works with AI into reusable skills. Find the context they keep supplying, the corrections they keep making, and the processes worth preserving. Aim for useful first outputs with less repeated explanation.
Use actual conversation exchanges as primary evidence. Memory can help locate relevant work, but is not a substitute for reading drafts, feedback, and revisions. Do not assume unstated preferences or invent a workflow from a job title.
Read the history that is actually available
Check the current conversation, supplied transcripts and files, and any available tools that can list, search, or read the person's other chat sessions. Use those tools when available. A skill does not create access to history, files, or attachments that the environment cannot retrieve.
Discover the range of accessible work before focusing on recurring workflows. Read the relevant exchanges, including revisions and available final artifacts. Summaries and search snippets can locate sessions; retrieve the underlying messages before treating them as evidence for detailed instructions. Continue through available pages where needed. Do not claim a complete scan when retrieval, context, pagination, or access limits prevented one.
Briefly tell the person what you examined and any meaningful gaps. Keep coverage separate from findings: distinguish fully read conversations from excerpts, summaries, inaccessible attachments, and unexamined sessions. Cite only material actually inspected.
If history is insufficient or substantially restricted, explain the specific limit and offer exactly two paths:
- Continue with what is available. Produce only what that evidence supports, identifying what remains unknown.
- Add more context. Supply relevant conversations, revisions, or final examples. If they have no history, offer to discuss one recurring task and the outcome they want.
Wait for their choice. Do not repeatedly ask about a limit they already accepted. With no usable evidence, do not fabricate history-based recommendations; explain this within the first option. If they choose to describe a task, ask the next useful question rather than a long intake questionnaire, and label the resulting requirements as newly supplied.
Extract what should become reusable
For each promising workflow, connect the original request, supplied context, draft, user corrections, subsequent changes, and any supplied or explicitly accepted final result. Look for:
- Recurring tasks and the sequence of work actually followed.
- Context, inputs, constraints, wording, formatting, and quality requirements the person repeatedly adds.
- Explicit desired outcomes and what the person says makes an output useful.
- Existing skills or resources the person explicitly asks to use.
- Requirements that could be included at the start instead of rediscovered through the same corrections.
Preserve the scope of each instruction. A preference for one client, document type, or situation is not automatically a rule for all work. A single explicit standing instruction can be sufficient evidence; repetition alone does not make a one-off request universal.
Do not treat an assistant's suggestion, an unrevised draft, or a conversation ending as user approval. Inspect a supplied final artifact when accessible; otherwise state that its contents could not be checked. If versions or instructions conflict and the history does not explicitly resolve them, ask a focused question before encoding the disputed rule.
Keep observed instructions, uncertain interpretations, and proposed improvements distinct. Clarify an inference that would change the generated skill. For example: "You asked for short summaries in these client updates. Should that apply to internal reports too?" Do not make the person re-answer things already clear in the history.
Prioritize well-supported recurring work where a reusable method would avoid repeated context or corrections. Do not invent time-saving estimates or force every repeated request into a skill.
Recommend a structure and let the person choose
Show a short set of findings with source references: the recurring work, what the skill would remember or do, and why it is worth creating. Use actual workflow names from the evidence.
Recommend the fewest skills that preserve meaningful differences in purpose, process, inputs, and quality standards. Related jobs can share a broader skill with situation-specific reference files. Split workflows when combining them would create confusing triggers, incompatible instructions, or unrelated processes. Shared wording preferences alone do not make unrelated workflows one skill.
Ask this packaging question, filling in the actual names and groupings:
- Separate skills: individual skills for the distinct workflows shown.
- One combined skill: one entrypoint with supporting references for the different workflows.
- Recommended grouping: the smallest sensible set, stating the proposed skills, what each covers, and a brief reason.
- Other: the person's preferred arrangement.
Wait for the choice; do not silently select the recommendation. If they already specified an arrangement, follow it without asking again. Do not invent extra workflows just to make the options different.
Ask before changing the process
Build clearly established requirements into the chosen skill. Separately propose a process change only when it would meaningfully help. Explain it briefly and ask whether to use it or keep the existing process.
For example: "You usually add audience details after the first draft. This skill could ask for them upfront to reduce revisions. Use that approach, or keep your current process?"
If accepted, incorporate the change. If declined, preserve the existing process. If unanswered, leave that change pending rather than treating silence as approval. Keep questions grounded in the inspected work and ask only what affects the result.
Create the selected skills
After resolving the choices and material gaps, read references/creation-and-evidence.md and create the actual skill files and useful supporting references. Do not stop at a list of ideas when the person has chosen what to create.
Deliver a concise explanation of the created skills, their intended outcomes, source support, and how to start using them in the verified target environment. Distinguish created files from installed or tested skills. Keep personal source material in the person's output; publishing this discovery skill does not authorize publishing their conversations or generated personal skills.