A handoff that writes to a file before /clear, /prime to resume, a mid-work checkpoint and a memory audit. Each does one job, in PT and EN.

skillmanager3x is a set of skills (ready-made instructions in text files) for people who use Claude Code and work in long sessions. They solve the problem of losing the thread when the conversation fills up: session-handoff writes the state to a file, you clear with /clear, and prime resumes by reading that file. There is also a quick checkpoint, a memory audit and a status bar. To use them, just copy the folders into ~/.claude/skills.
There is no build, no dependencies and no executable code. Each skill is a folder with a SKILL.md (markdown + YAML frontmatter) that Claude Code loads when you say one of the triggers.
Passes the baton before /clear or an agent switch. Shows it in the chat and writes it to handoffs/history/<UTC>.md + handoffs/latest.md, with the exact next action, real verification and a secrets check.
The handoff's partner: at the start of the session it reads CLAUDE.md/AGENTS.md, context/, tasks/ and handoffs/latest.md and returns a short briefing. Read-only; it doesn't resume a deploy or a send on its own.
A quick checkpoint during the work: current focus, what went to memory, recent decisions, next steps and a recommendation (continue, compact, audit, handoff or clear).
An inventory of what is in ~/.claude/projects/<project>/memory/ and in CLAUDE.md. It separates durable facts (keep) from short-term state (discard) and noise (remove).
Rule of thumb: end of session โ handoff. "Where are we?" โ statusline. "What's in memory?" โ audit. You speak in natural language, the harness matches the frontmatter description and loads the right skill.
"handoff", "wrap up", "wrap up session", "summarize before clear", "I'm going to /clear".
"prime", "resume", "pick up where we left off", "read the handoff", or when entering a project that has handoffs/latest.md.
"statusline", "checkpoint", "where are we", "next steps".
"audit memory", "memory audit", "what was saved", "bloated memory".
When the context fills up or the session has already compacted, don't insist on the same conversation. Write the state to a file, clear, and start over by reading that file. This is the process we recommend for any long work, in Claude Code and in Codex.
Writes what was done, what is in progress, the exact next action and what depends on the human, with no secrets. It writes a new file to handoffs/history/<UTC-date>.md (never overwrites) and updates handoffs/latest.md.
Starts from scratch: empty context, without the weight of compactions. Nothing is lost, because the state is in the file. In Codex, open a new session.
Before acting, it reads AGENTS.md/CLAUDE.md, context/, tasks/current.md and handoffs/latest.md and returns a short briefing: goal, main rule with source, last decision, next action and conflicts. Read-only: it doesn't resume publishing, deploy or sending on its own.
Old version (chat only), the one this repository shipped until Oct/2026 and that circulates in other kits: it shows the summary in the conversation and writes nothing. When you run /clear, the summary goes away with it unless you copy it by hand, and /prime finds nothing to read.
Current version (writes to a file), the one now in skill/pt and skill/en, identical to the one in the agente-claude-codex 1.5.2 kit: it shows in the chat and writes to handoffs/ through the handoff-grava.py script, with latest.md as an index of open work (the same format the Long Execution hook reads); it creates the folder if the project only has CLAUDE.md.
# 0 = chat-only version; 1 or more = writes to handoffs/ grep -c "handoffs/history" ~/.claude/skills/session-handoff/SKILL.md # swap in the version that writes (and install the prime partner) cp -r skill/en/session-handoff skill/en/prime ~/.claude/skills/
This cycle is the foundation of the Long Execution project, the method for agents that work for hours or days (/goal, state files, ceilings and a done criterion). It has the hook that measures the context and writes the handoff on its own at ~70% and ~85% and right after a compaction, while the cache is still warm; you only run /clear and /prime. It also has the research behind the context bands and recall to search everything that has ever been said.
If you already use Claude Code, you're ready. There's no package to install, no runtime and no service to run.
Anthropic's CLI, with skills support (the skills/ folder is scanned at start).
# check that it's installed claude --version
Clone (or download) it โ the skills are just text files.
git clone https://github.com/inematds/skillmanager3x cd skillmanager3x
Global (all sessions) or per project. Create it if it doesn't exist.
# global mkdir -p ~/.claude/skills
Five steps. The longest takes a minute.
The skills live in skill/pt/ and skill/en/ โ same structure in both languages.
git clone https://github.com/inematds/skillmanager3x cd skillmanager3x
The name and description fields stay in English in both versions (for consistent harness triggering), but the description in pt/ includes Portuguese triggers.
ls skill/pt # session-handoff prime session-statusline memory-audit ls skill/en # same structure, texts in English
Copy the whole folder of each skill โ with the SKILL.md inside. Don't flatten the structure.
# global โ applies to all sessions mkdir -p ~/.claude/skills cp -r skill/en/session-handoff ~/.claude/skills/ cp -r skill/en/prime ~/.claude/skills/ cp -r skill/en/session-statusline ~/.claude/skills/ cp -r skill/en/memory-audit ~/.claude/skills/ # or per project cp -r skill/en/session-handoff <project>/.claude/skills/
In Claude Code, list the available skills. All four should appear.
/skills # session-handoff ยท prime ยท session-statusline ยท memory-audit
You don't need to memorize commands โ speak naturally and the right skill kicks in.
# end of session: writes handoffs/latest.md, then /clear handoff before I /clear # new session: reads the handoff and summarizes /prime # mid-work statusline # review memory audit this project's memory, it's messy
Three formats, three purposes. Everything comes out in the chat; only the handoff also writes to handoffs/, which is where prime reads from.
Written to handoffs/history/2026-10-13T014500Z.md and copied to handoffs/latest.md.
# Handoff โ 2026-10-13 โ auth middleware refactor ## Accepted state - [confirmed] auth/jwt_validator.py extracted from middleware.py:45-180 ## Proposed / tried, not confirmed - [unverified] public-key cache TTL 5min โ no test - [?] v1โv2 migration โ user wanted to review it first ## Verification - Run: `pytest tests/auth` โ 12 passed ## Running state - Background processes: shell_a3f โ pytest --watch ## Exact next action Check whether the v1โv2 migration goes in now (needs the user's ok).
It always ends with a Recommendation: which skill (or none) makes sense next.
# Statusline ## State - Current focus: adjust pagination on the /users endpoint - Context: moderate - Risk: no load test with >10k records ## What was saved to memory - Confirmed saved: none - Candidates: prefers keyset over offset ## Next steps 1. Run a test with 50k records 2. Update the OpenAPI schema ## Recommendation continue
The output is a recommendation, not an execution. If you want to apply the cuts, ask in a second step: "ok, apply that".
# Memory Audit ## Keep (durable facts) - user_role.md โ user is a Go backend engineer, new to React - feedback_testing.md โ integration tests on real Postgres (Q3/2025 incident) ## Consider moving to CLAUDE.md - project_auth_rewrite.md โ motivation (compliance) โ belongs in the project docs ## Remove (short-term / noise) - session_2026-04-15.md โ old session state - bug_fix_recipe_xyz.md โ fix already applied, the code speaks for itself ## Points of attention - Duplication: feedback_pr_size.md and feedback_review_style.md - Stale: reference_grafana.md points to a dashboard whose URL changed
Besides the three skills, the repo ships a statusline for Claude Code: the fixed line below the input, with directory, model, effort, remaining context and how much of your limit is already gone. One command to install.
# what shows up in the terminal nei@maquina:~/projetos/wifi Opus 5/low ctx 10% 5h:24%โ00:10 w:29%โ10/08 21:00 d:109.6M
user@host:cwd plus Opus 5/low โ the session's model and effort level, following what you chose in /model.
ctx 10% โ how much of the window has been used. It's the number that tells you whether you can keep going or it's time for a handoff.
5h:24%โ00:10 and w:29%โ10/08 21:00 โ usage and reset time of the 5-hour and weekly windows. d:109.6M is the day's tokens.
The script copies the statusline to ~/.claude/ and registers the statusLine block in your settings.json โ preserving the rest of the file, with a .bak backup.
git clone https://github.com/inematds/skillmanager3x bash skillmanager3x/statusline-cmd/install.sh
The installer prints the line already assembled with your data. It shows up on Claude Code's next render โ no session needs to be restarted.
# dependencies, if any is missing sudo apt install jq curl python3
The uninstaller removes only the statusLine key from settings.json; the rest of your settings stays intact.
bash statusline-cmd/install.sh --uninstall
The 5h and weekly percentages come from Anthropic's usage endpoint using your own credentials (~/.claude/.credentials.json) โ no token goes in the repository. Details, available fields and how to customize: statusline-cmd/README.md (in Portuguese).
Similar names, different things: the session-statusline skill writes a text checkpoint when you ask; the statusline-cmd folder installs the fixed terminal bar.
The project is small on purpose: disjoint skills, no trigger overlap.
SKILL.md in both languages and detailed READMEs in skill/en/ and skill/pt/ covering design decisions.session-handoff now writes to handoffs/history/ + handoffs/latest.md (same version as the agente-claude-codex kit, now 1.5.2) and prime joined, which reads that file in the new session. Full cycle: handoff โ /clear โ /prime./goal for work lasting hours: Long Execution guide.