Human context rebuild
Rebuild the user's mental model of the current session with a tight recap — what we're doing, why, and where we are. Use this to recover after a break, a tangent, or a long subagent run. Trigger when the user invokes /human-context-rebuild, or says "remind me what we're doing", "我忘了", "我们到哪一步了", "where are we", "what was I doing", "summarize this session", "我们在干啥来着", "这是啥来着", "这个是干嘛的来着", "想不起来这个 session 在做什么", "脑子糊了", or similar re-orientation cues — including a bare "?" or "这是啥" pointing at the current work. Default to firing when the user signals lost context about the session itself; the cost of an unneeded recap is low. Do NOT trigger for "explain this code", "summarize this file", or any task-summary request — those are different.From its SKILL.md
npx -y skills add lycfyi/yskills --skill human-context-rebuildAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 22 stars22 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its file declares
Copied from the file, not written here
The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.
SKILL.md
3.6 KB, as published. Nobody here has run it
Human Context Rebuild
Purpose
Drop the user back into the session in under 30 seconds of reading. They forgot. Their job is to scan and resume — not to re-read a transcript.
Produce one short recap. No preamble, no "great question", no offering to do more.
Output structure
Use exactly these sections, in this order. Skip a section only if there's genuinely nothing to put in it (don't pad).
**现在在做 / Now**: <1 line — the active task in concrete terms>
**为什么 / Why**: <1 line — the trigger or motivation the user gave>
**目标 / Goal**: <1 line — the success condition>
**进展 / Done**: <bullet list, ≤4 items, most recent first>
**下一步 / Next**: <1–2 lines — the very next action or open decision>
**Open questions** (only if blocking): <bullets>
Total length: aim for under 150 words. Hard cap 250.
Language
Match the session's dominant language. If the user has been mixing Chinese and English, mix in the recap too — keep technical terms (file paths, function names, statuses) in their original form. Don't translate identifiers.
What to include
- Concrete anchors: file paths, function names, branch names, decisions the user explicitly confirmed. These are what the user's brain latches onto faster than prose.
- Decisions made, especially scope changes ("user dropped quota, kept only AI judge").
- The most recent meaningful action: what the last tool call or agent run actually produced.
What to leave out
- Tool-call mechanics (which agent ran, how many files were grep'd).
- Detailed code or SQL — link to the file path instead.
- Praise, hedges, "let me know if…", or offers to continue. The user can see the next step; they'll drive.
- Things the user already said in this turn. Don't echo their own message back.
- Re-justification of past decisions. State them as settled.
Edge cases
- Very new session (under ~3 turns): say so directly. Example:
我们才刚开始 — 你给的方向是 X, 还没动手. 要不要先 ___? - Session changed topics mid-way: recap the current topic only. Mention the prior topic in one line if it's still open:
(早些时候在弄 Y, 暂停了.) - Just finished a big chunk of work: lead Progress with what shipped, then Next with verification or the follow-up commit.
- In the middle of a blocked tool call or pending question: put the blocker in Open questions and make Next be "answer the question above".
Self-check before sending
- Could the user paste this into a new conversation and pick up the thread? If not, add the missing anchor.
- Is anything in the recap something they explicitly said this turn? Cut it.
- Over 250 words? Cut Progress bullets first, then trim Why.