| name | completion-learn |
| description | Use when the user explicitly asks for /learn-complete or a post-task debrief after work is finished. Extract what should remain after a long task across self, collaboration, and tool. Never auto-trigger mid-task. |
Completion Learn
Overview
Use this skill for an explicit completion debrief. The goal is to extract what should remain after a finished task across self, collaboration, and tool.
If the user only says /learn-complete, interpret it as "help me see what should remain after this task."
Default mode is three-axis evolution:
Prefer a short debrief with durable residue over abstract richness.
Treat task completed and capability gained as separate judgments.
Rules
- Require explicit invocation. Never slide into this mode during normal execution or mid-task updates.
- Only run when the task is complete enough for an honest retrospective. If completion is unclear, say so and stop.
- If the user gives a focus like
debugging, workflow, or collaboration, treat it as the primary lens.
- Prefer the smallest useful output shape from
references/output-structure.md.
- Focus on what should remain from the long interaction instead of replaying the session.
- Before writing the axes, privately judge the residue level:
mostly assisted performance
partial internalization
durable capability gain
- Keep that residue judgment internal by default. Do not add a fourth visible section unless the user asks for deeper explanation or the distinction is the main lesson.
- Inside
Self, inspect the user's side of the interaction: attention, phrasing, emotion, ownership, and judgment habits.
- Inside
Collaboration, inspect what interaction pattern between the user and AI should become default.
- Map mistakes to one or two labels from
references/mistake-patterns.md.
- In default mode, keep the priority order
Self > Collaboration > Tool.
- End with one durable rule and one next deliberate practice.
- Make
next deliberate practice a reduced-support test, not generic advice. It should check whether the claimed residue actually remains when one layer of help is removed.
- If the residue is
mostly assisted performance, keep the practice narrow and ask the user to do one small step alone before using the tool.
- If the residue is
partial internalization, make the practice user-first and tool-second on the same bottleneck.
- If the residue is
durable capability gain, move the practice to a nearby transfer task instead of repeating the exact same one.
- Inside the
Tool axis, keep skill sedimentation separate from skill evolution / optimize.
- If the current skill boundary is right but weak, route to
skill-optimizer.
- If the boundary itself should change or a second complementary skill is needed, route to
skill-creator.
- If confidence is low, stop at recommendation instead of chaining more skill work.
When to Use
Use when:
- the user invokes
/learn-complete
- the user asks for a post-task retrospective, completion debrief, or "what should remain from this task?"
- the user wants to know whether the finished workflow should optimize an existing skill or become a new one
Do not use when:
- the task is still underway
- the user wants a checkpoint, status update, or implementation work
- the main need is debugging, editing, or normal summarization without reflection
Typical asks:
/learn-complete
/learn-complete debugging
顺便看一下,这次该优化已有 skill 还是新建一个
Workflow
- Confirm completion and pick the primary lens.
- Choose
Light, Standard, or Deep from references/output-structure.md.
- Judge whether the outcome was mostly assisted, partially internalized, or a durable capability gain.
- Extract the strongest residue worth keeping, the main mistake pattern, and the missing guardrail.
- Convert that into
Self, Collaboration, and Tool sedimentation without changing the visible three-axis shape.
- End with one durable rule, one next deliberate practice, and one tool-evolution judgment.
- Make sure the practice step matches the residue judgment by reducing support or forcing transfer.
- Route to
skill-optimizer, skill-creator, or stop at recommendation.
Resource Map
references/improvement-axes.md, references/self-improvement.md, references/collaboration-evolution.md -> default three-axis rubric
references/focus-lenses.md, references/question-bank.md -> choose the learning lens and deepen it
references/mistake-patterns.md, references/output-structure.md, references/sample-output.md -> shape the output and keep it concrete
references/skill-sedimentation.md, references/skill-evolution.md, references/tool-evolution-routing.md -> decide whether to optimize, extend, add, or stop
references/examples.md -> representative user phrasings