observe
Memory observer agent for recording structured observations from primary Claude Code sessions.
Skill by yiliu30
Chalk Labs
Need help deploying observe?
Chalk Labs sets up agents like this inside real teams: connected to your tools, tested on your data, with a human in the loop where it matters.
Install
Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.
npx agentmarket install skill/yiliu30-observe
Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.
Get the weekly diff
The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.
About
Use cases: writing, productivity, agent-building
Reviews
Write a reviewNo reviews yet. Installed it? Reviews come from verified installers only. How reviews work
Permissions
What this skill can do and what it asks for, from its SKILL.md. Checked Oct 25, 2026.
- Tools
- No tools found
- Secrets
- None asked for
- Install scripts
- None found
Network hosts it talks to 1
codeload.github.com· found in install command
Files it touches
~/.claude/skills/observe
What's inside
- Skill name
observe- Description
- Memory observer agent for recording structured observations from primary Claude Code sessions. Use this skill whenever you're asked to observe, monitor, or record what's happening in another session. It prevents low-yield observation sessions by enforcing a "bail early" principle — if there's nothing meaningful to observe, write a short note and end the session instead of lingering.
- License
- MIT
- Instructions
- 74 lines in SKILL.md · view on GitHub
Files (1)
- SKILL.md
SKILL.md (first lines)
# Memory Observer You are monitoring a primary Claude Code session and recording observations for future session continuity. ## First: Decide Whether to Observe or Bail Before writing anything, assess the primary session context. This is the most important step — the #1 source of wasted observer sessions is lingering when there's nothing to observe. **Bail immediately** (write a minimal note per Step 2, then end your session) if: - The primary session is just polling, sleeping, or waiting for long-running jobs - No primary session context was provided - You've already written an observation for this session and nothing new has happened - The only visible activity is repeated status checks with no change **Observe** (write a full observation per Step 3) if: - Code was explored, read, or modified - Technical findings were made (even partial ones) - Decisions were made or errors were debugged - New files or artifacts were produced If it's genuinely ambiguous, write a condensed observation covering only what you can cite with evidence, and explicitly note what's missing. ## Step 1: Check for Existing Observations Before recording, check if `observations/` already has a recent file for the same session or topic. If a prior observation already covers the same ground, only record **new** findings. Don't duplicate. ```bash [ -d observations ] && ls observations/ | tail -5 ``` ## Step 2: Minimal Note (bail case) When there's nothing meaningful to observe: ```markdown # Session Note — YYYY-MM-DD HH:MM **Status:** No meaningful activity observed **Primary session:** [what was visible, e.g., "polling video generation progress every 90 min"] **Duration monitored:** [approximate time] **Recommendation:** [when to check back or what to wait for] ``` Write this, save it (using the format in Step 4), and **end your session.** Don't pad it or keep watching. ## Step 3: Full Observation (observe case) Write a structured observation. Two rules: 1. **Every claim needs evidence.** Cite file paths with line numbers, command outputs, or document references. "Investigated CUTLASS source" is useless. "Found FP4 accumulator constraint in `mma_traits_sm100.hpp:247`" is useful. 2. **Skip sections you can't fill.** Only include sections where you have real content. An observation with 2 strong findings beats one with 5 sections of filler. Sections to use (skip any that are empty): - **Context** — session goal, project, whether context was rich or partial - **Key Findings** — with evidence citations - **Decisions Made** — with rationale - **Problems Encountered** — what failed and what worked (future sessions need to know what NOT to retry) - **Artifacts Produced** — files created/modified, PRs, issues - **Open Questions** — unresolved items for next session - **Next Steps** — concrete actions for the next session to pick up Always include Next Steps if there's any ongoing work — future sessions rely on this to know where to start. ## Step 4: Save Save to the project's observation directory: ``` observations/YYYY-MM-DD-HHMM-[brief-slug].md ``` Create `observations/` if it doesn't exist. Use descriptive slugs: `mxfp4-accumulator-findings`, not `session-notes`.
From yiliu30/z2h, MIT.
More from yiliu30/z2h
- 0.563score
profile-cute-xpu
yiliu30
Profile and analyze CuTe/SYCL*TLA kernels on Intel XPU/GPU, especially Intel Xe/BMG/Xe20/Xe35 GEMM-style kernels using CuTe primitives such as XE_DPAS_TT…
SkillBFreshness: B★ 2Claude Code +9 - 0.559score
intel-xpu-cute-accuracy-debug
yiliu30
Debug numerical mismatches in Intel XPU CuTe/CUTLASS kernels, especially Xe DPAS/BDPAS fragment layouts, block-scale payloads, and fused SDPA/GEMM accumulators.
SkillBFreshness: B★ 2Claude Code +9 - 0.507score
remote-debug
yiliu30
Drive Claude Code to debug an issue on a remote node over SSH — connect, set up an environment, copy scripts/probes across, run repros, and capture findings.
SkillCFreshness: C★ 2Claude Code +9 - 0.507score
confluence-html
yiliu30
Create rich, interactive HTML that survives Confluence's save sanitizer by using only inline styles, no external resources, and JavaScript for dynamic content.
SkillCFreshness: C★ 2Claude Code +9 - 0.501score
read-arxiv-paper
yiliu30
Given an arxiv URL, download the TeX source, read the paper, and produce a summary markdown file
SkillCFreshness: C★ 2Claude Code +9 - 0.448score
reduce-model
yiliu30
Create a reduced (fewer-layer) version of a large HuggingFace model for fast testing and debugging.
SkillDFreshness: D★ 2Claude Code +9 - 0.443score
find-sth
yiliu30
Investigate a technical question by searching across multiple sources (source code, documentation, academic papers, existing findings) and produce a sourced…
SkillDFreshness: D★ 2Claude Code +9 - 0.443score
real-review
yiliu30
Fact-check written content (GitHub issues, docs, reports, posts) by verifying every claim, number, link, code reference, and conclusion against actual evidence.
SkillDFreshness: D★ 2Claude Code +9 - 0.443score
vllm-kernel-trace
yiliu30
Trace which GPU kernels vLLM dispatches for specific model layers (linear, attention, MoE, norms).
SkillDFreshness: D★ 2Claude Code +9
Signals
- Stars
- 2
- Forks
- 0
- Open issues
- 1
- Last commit (this folder)
- Apr 12, 2026 (6 months ago)
From GitHub (fetched Oct 24, 2026).
Details
- License
- MIT
- Repository
- github.com/yiliu30/z2h
- First seen
- Oct 24, 2026
- Last checked
- Oct 24, 2026
- Source
- GitHubgithub-graphql
More from yiliu30
- 0.563score
profile-cute-xpu
yiliu30
Profile and analyze CuTe/SYCL*TLA kernels on Intel XPU/GPU, especially Intel Xe/BMG/Xe20/Xe35 GEMM-style kernels using CuTe primitives such as XE_DPAS_TT…
SkillBFreshness: B★ 2Claude Code +9 - 0.559score
intel-xpu-cute-accuracy-debug
yiliu30
Debug numerical mismatches in Intel XPU CuTe/CUTLASS kernels, especially Xe DPAS/BDPAS fragment layouts, block-scale payloads, and fused SDPA/GEMM accumulators.
SkillBFreshness: B★ 2Claude Code +9 - 0.507score
remote-debug
yiliu30
Drive Claude Code to debug an issue on a remote node over SSH — connect, set up an environment, copy scripts/probes across, run repros, and capture findings.
SkillCFreshness: C★ 2Claude Code +9 - 0.507score
confluence-html
yiliu30
Create rich, interactive HTML that survives Confluence's save sanitizer by using only inline styles, no external resources, and JavaScript for dynamic content.
SkillCFreshness: C★ 2Claude Code +9 - 0.501score
read-arxiv-paper
yiliu30
Given an arxiv URL, download the TeX source, read the paper, and produce a summary markdown file
SkillCFreshness: C★ 2Claude Code +9 - 0.448score
reduce-model
yiliu30
Create a reduced (fewer-layer) version of a large HuggingFace model for fast testing and debugging.
SkillDFreshness: D★ 2Claude Code +9
Done for you
Rather not set it up yourself?
Chalk Labs scopes, builds and runs agents for your team: the task audit, the eval harness, the integrations and the weekly ops report. Fixed scope, priced before we start.
See how we build AI agents →Read the case studies
Chalk Labs is a two-founder AI and growth studio. We build AI agents, CRMs and MVPs for Web3 and AI companies, and have shipped 50+ product and token launches.
- launches shipped
- 50+
- launches shipped
- media placements
- 500+
- media placements
- founders on every account
- 2
- founders on every account
Press we've placed: Forbes · CoinDesk · Decrypt · Cointelegraph · The Block
Rahil Jain, Co-founder, Chalk Labs. Ex-Goldman Sachs engineer. You talk to the person who builds it.
Get observe running in your business
Tell us what you want to automate. A founder replies within one working day with a plan, not a pitch deck.
Cite this page
observe skill. Agent Mart. https://agentmart.dev/skills/yiliu30-observe (last updated 2026-10-24)