Built by Andy Toizer — I'm the head of growth at Freckle.io and write AgentOperator, a newsletter about what it actually looks like to build real systems with coding agents as a non-engineer.
TLDR: A Claude Code skill that keeps context tight across long projects — either by compacting within a session or handing off cleanly between sessions. It maintains a persistent README as the source of truth so no work gets lost.
It came from running multi-session outbound campaigns, data pipeline builds, and tooling projects where context kept getting lost between conversations. After rebuilding the same context three times, I turned the pattern into a skill. Later, I realized most "session handoffs" should actually be compactions — you only need a new session when the task genuinely changes.
When a project takes more than one Claude Code session, every new session starts cold. You re-explain context, re-make decisions, and waste time getting back to where you were. But starting a new session isn't always the right move — sometimes you just need to compress what you have and keep going.
flowchart TD
A["Big project lands"] --> B{"Will this take\nmore than one session?"}
B -->|No| C["Just do it"]
B -->|Yes| D["Create project folder\nwith README.md"]
D --> E["Execute current phase"]
E --> F{"Context getting heavy?"}
F -->|No| E
F -->|Yes| G{"Does the next step need\nwhat's in my head,\nor just what's in the files?"}
G -->|"In my head\n(iterating, refining,\nuser feedback not yet captured)"| H["Compact"]
G -->|"In the files\n(new task, different tools,\nREADME has everything)"| I["Session Handoff"]
H --> J["Update README"]
J --> K["Compress context in place"]
K --> E
I --> L["Update README"]
L --> M["Generate handoff prompt"]
M --> N["New session reads README\nand picks up immediately"]
N --> E
style D fill:#ffe0b2
style J fill:#ffe0b2
style L fill:#ffe0b2
style G fill:#e1f5fe
The skill does four things:
-
Proactive splitting — When you bring a big project, it assesses the scope and proposes session boundaries before starting.
-
Persistent state — Creates a project folder with a README.md that tracks everything: plan, status, decisions, and a session log. This is the single source of truth.
-
Smart context management — When context gets heavy, it decides whether to compact (preserve nuance, keep working) or hand off (clean break, fresh start). The decision framework is simple: if a colleague could pick up from the README alone, it's a new session. If they'd need to ask "what did the user say about X?", compact instead.
-
Clean handoffs — When a new session is the right call, it generates a ready-to-paste prompt so the next session reads the README and hits the ground running.
| Compact | New Session |
|---|---|
| Iterating on the same work (refining, re-running with tweaks) | Starting a genuinely different task with different tools |
| User gave feedback you haven't captured in files yet | Everything needed is in the README + data files |
| Mid-loop — not at a natural stopping point | At a clear phase boundary |
| Tool state and working context matter | Current context would just be noise |
| Next step is a continuation | Next step has its own setup and iteration loop |
Rule of thumb: If you could hand the README to a colleague and they'd know exactly what to do without asking you questions, it's a new session. If they'd need to ask "wait, what did the user say about X?" — compact.
# EdTech Outbound Campaign
## Overview
Targeted outbound to 200 companies for Freckle...
## Status Tracker
### Phase 1: Build Company Lists ← COMPLETED
- [x] Apollo org searches
- [x] AI Ark API client built and tested
### Phase 2: Find People ← CURRENT
- [x] Apollo search — 1,689 contacts across 157 domains
- [x] AI Ark enrichment — 887 matched
- [ ] Email enrichment waterfall
## Decisions Made
1. Hybrid approach: Apollo → AI Ark → Web Search
2. C-suite deprioritized at 500+ employee companies
3. Product marketing excluded from marketing leader segment
## Session Log
### Session 1 — 2026-04-15
**Completed:** Built AI Ark client, collected 300 raw companies
**Next:** Curate to 100 per vertical using AI Ark lookalike searchOpen a new chat and paste this:
install https://github.com/Andytoizer/session-relay
Claude will fetch skill/SKILL.md from the repo and install it to your skills directory. Confirm when prompted. After that, the skill is active in every session — you don't need to reference it explicitly.
# Clone the repo
git clone https://github.com/Andytoizer/session-relay
cd session-relay
# Copy to your skills directory
mkdir -p ~/.claude/skills/session-relay
cp skill/SKILL.md ~/.claude/skills/session-relay/SKILL.mdmkdir -p ~/.claude/skills/session-relay
curl -o ~/.claude/skills/session-relay/SKILL.md \
https://raw.githubusercontent.com/Andytoizer/session-relay/main/skill/SKILL.mdThat's it. Claude Code will automatically use the skill when it detects a multi-session project.
The skill activates when:
- You start a project that will clearly take more than one session
- Context is getting long on a complex task
- You say things like "let's split this up" or "we'll continue next time"
- You've completed a phase and the next one is genuinely different work
It also triggers proactively — if Claude recognizes the scope is too big for one session, it'll propose a plan without you asking. And when context gets heavy mid-session, it'll suggest compacting rather than forcing a session break.
Good:
Completed: Built and tested the API client. People search works with domain + seniority filters. Raw data collected: 156 EdTech + 166 ConstructionTech companies.
Next session should:
- Read
campaigns/README.mdfor full context- Curate EdTech list to 100 using lookalike search seeded from ClassDojo
- Filter ConstructionTech — Apollo returned actual construction firms, not software companies
Bad:
We made good progress. Next time we should keep working on it.
The difference is specificity. File paths, numbers, exact next steps.
MIT — use it, fork it, improve it. If you build something cool, let me know.