菜单
Extract L1 contributor atoms from a GitHub subject's raw activity. Triggers after "tmem contrib raw <id>" or when the user says "ingest contributor", "profile <user>", "analyze how <user> works". This CREATES atoms from raw GitHub events — for synthesis use contrib-synthesize.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | contrib-ingest |
| description | Extract L1 contributor atoms from a GitHub subject's raw activity. Triggers after "tmem contrib raw <id>" or when the user says "ingest contributor", "profile <user>", "analyze how <user> works". This CREATES atoms from raw GitHub events — for synthesis use contrib-synthesize. |
Turn one subject's raw GitHub activity into evidence-linked L1 atoms across the 11 fixed dimensions. You do all classification — no external LLM.
tmem contrib raw <subject-id>
This prints {commits, prs, reviewCommentsGiven, reviewThreadsReceived, issues}
(bots/forks/generated files already filtered):
commits — the subject's commits across ALL branches (default branch + every
PR's head branch), deduped by sha.prs — PRs the subject authored (all branches, via the search API).reviewCommentsGiven — review comments the subject WROTE on others' PRs.reviewThreadsReceived — comments on the subject's own PRs (with is_subject
flagging their own replies vs reviewers').If it errors with "gh not found" or auth failure, tell the user to run
gh auth login and stop.
Before classifying, read references/dimensions.md — the per-dimension
rubric with good-vs-shallow atom examples, the Ousterhout lens for solve, and
evidence-strength criteria. Classification quality depends on it; the summary
below is only the map.
For each meaningful signal, write ONE atom tagged with exactly one dimension.
Never invent style — every atom needs at least one evidence link (PR#<n> or
commit sha).
Technical Craft
idea — how they frame problems / pick work. Source: issue bodies (repro,
expected-vs-actual, root-cause vs symptom), PR "why" sections.plan — PR decomposition & scoping. Source: PR size (additions+deletions),
commits-per-PR, whether each PR is self-contained.solve — coding/refactor patterns. Read diffs through Ousterhout's lens: deep
vs shallow modules, information leakage, strategic vs tactical, errors designed
out of existence.craft — review thinking in reviewCommentsGiven: do they cite the why,
weigh alternatives, label severity ("Nit:", "Optional:").Collaboration & Influence
comms — commit message quality (subject ≤50 chars, body explains why,
imperative mood) and PR description clarity.mentor — reviewCommentsGiven that teach/explain vs cosmetic-trivia floods.conflict — reviewThreadsReceived: in their own replies (is_subject:true)
do they update their view, push back constructively, avoid needless blocking.Outcomes & Ownership
scope — cross-repo/cross-area reach, size of areas touched.ownership — test-inclusion rate, concentration on components.execution — revert rate, post-merge rework, review coverage of merged work.tmem contrib upsert-atom --json '{"record_id":"<id>:<dim>:<hash>","subject_id":"<id>","dimension":"plan","content":"Splits features by concern; median PR ~280 LOC, ~5 commits each.","evidence":["PR#1234","PR#1240"]}'
record_id must be stable (e.g. <subject-id>:plan:<short-hash-of-claim>) so
re-ingest upserts instead of duplicating.content to one concrete, emulable observation.Tell the user how many atoms were written per dimension and any dimensions left empty for lack of evidence.
Use the tmem CLI to inspect and manage the local memory store — record/vector counts, keyword/hybrid search, view persona, list scenes, read one full scene block (tmem scene <name>), recent changes, sync vectors, configure thresholds. Also covers CROSS-PROJECT exploration — search every project store at once (tmem search <q> --all), list all memory stores (tmem projects), and collapse legacy cwd-keyed fragment stores into their project root (tmem migrate-fragments). Trigger when the user asks "how many memories do I have", "show my persona", "what scenes exist", "search memories for X", "search across all projects", "what projects/stores do I have", "my memory is fragmented", "open that scene", "recent memory changes", or when you need to check memory state before/after an operation. Do NOT use for extracting or consolidating memories — those have dedicated skills.
Consolidate L1 memory atoms into L2 scene blocks and L3 persona. Triggers when the user says "consolidate memories", "build persona", "update persona", "update scenes", "organize memories", or after memory-seed completes. Also triggers via asyncRewake pipeline after N conversation turns. This skill is about ORGANIZING existing memories into higher structures — for creating memories from transcripts use memory-seed instead.
Extract L1 memory atoms from Claude Code conversation history. Triggers when the user says "seed memories", "extract memories", "backfill memory", "remember my history", "learn from past conversations", "what do you know about me", or after /memory-init when conversation history exists. Also use when asyncRewake pipeline flags pending sessions. This skill is about CREATING new memories from transcripts — for inspecting existing memories use tmem-cli instead.
One-shot entry point for Contributor Intelligence — the user just drops a GitHub link or names an engineer and you take it from there. Use this whenever someone pastes a GitHub profile/repo URL (github.com/<user> or github.com/<user>/<repo>) or a handle and asks to "profile", "analyze how X works", "learn from X", "study this engineer", "phân tích người này", "học từ người này", or asks "how do I use /contrib". Drives the whole pipeline end-to-end (add → ingest → build → playbook), or guides the user through it if they'd rather drive. This is the orchestrator — the per-phase skills (contrib-ingest, contrib-consolidate, contrib-synthesize) do the actual work; do NOT use this one to inspect or synthesize an already-built persona (use contrib-synthesize) or for the self-memory feature.
Consolidate a contributor's L1 atoms into an L3 persona across 11 dimensions. Triggers when the user says "build contributor persona", "consolidate <user>", or after contrib-ingest completes. This ORGANIZES existing atoms — for creating atoms use contrib-ingest.
Use this skill when someone wants to LEARN from engineers they've already profiled — extracting actionable lessons from contributor personas. Trigger on intents like: "what do the top SWEs / engineers / contributors I've profiled have in common", "extract their shared / common capabilities", "turn a contributor's profile into copyable habits, heuristics, or a learnable playbook", "what habits should I copy from <person>", or "compare me / my style to <role-model> dimension-by-dimension and find my gap". Also covers surfacing exemplar teaching quotes across profiled engineers. Operates on existing personas (built via contrib-consolidate); cross-engineer comparison or common-capability synthesis needs ≥2 subjects. Do NOT use for: building, ingesting, or inspecting a single persona's raw contents; commit/trajectory stats; summarizing documents; recalling what you know about the user; or playbooks unrelated to engineer profiles (e.g. team incident-response runbooks).