The Manifest · Issue 009 ·
After compaction, Claude Code keeps 5,000 tokens of each skill. 34 of ours are longer.
What a long skill loses mid-session and how Superpowers fixed it, a planner that keeps its plan on disk, and Agent Skills' move to the Linux Foundation.
Hi again.
This week: what Claude Code keeps of a skill when it compacts a long session, and the skill that lost a rule because of it; a planner built to survive that; the 34 skills in our catalog that run past the limit; a new plugin that reverse engineers shipped software; and Agent Skills’ new home at the Linux Foundation.
Same promise as always: what passed, what didn’t, and what we’re learning about which skills actually work. No generic AI news roundup. Just useful stuff.
THE SIGNAL / 01
When Claude Code compacts a long session, it re-attaches only the first 5,000 tokens of each skill you’ve used. Anything past that is dropped until you invoke the skill again.
Claude Code’s skills docs spell it out: after a compaction, it re-attaches the most recent invocation of each skill, “keeping the first 5,000 tokens of each.” All re-attached skills share a 25,000-token budget, filled from the most recently used skill backward, so in a session where you’ve used many, the oldest can be dropped entirely.
Superpowers found out what that costs. In a reported session, Subagent-Driven Development lost its five-round limit on fix attempts after a compaction (the rule sat past the cutoff) and ran 18 fix rounds on one task. Superpowers 7.0.1, released October 21, moves the fix loop, final review and the other late steps of that skill into references/ files that it reads from disk at the step that needs them. Writing Skills and Executing Plans were trimmed the same way, and a new test fails any SKILL.md that reaches 20,000 bytes, their stand-in for 5,000 tokens.
We measured our catalog against the same line: 34 of the 160 skills whose text we carry are longer, among them Claude API (about 105,000 characters), 15 of Microsoft’s Azure skills, Taste Skill and shadcn/ui. That isn’t a verdict on them; most are long because they carry reference material. It means item 1 below applies.
What to do with this:
1. If a long skill matters late in a session, call it again after a compaction. A fresh invocation loads it again.
2. If you write skills, put the rules that must hold at the top, and move reference material into separate files the skill reads when it needs them. That is what the docs recommend anyway (“Keep SKILL.md under 500 lines”).
3. Update Claude Code. Version 2.1.304 changed compaction to send Claude the skill list again afterward, so the agent still knows which skills it can reach for.
Also: Agent Skills has a new owner. On October 22, Anthropic gave the Agent Skills standard (the name, the GitHub organization and agentskills.io) to the Agentic AI Foundation at the Linux Foundation, where MCP already lives. Nothing about the SKILL.md format changed. Anthropic proposed versioned releases, a formal proposal process and security guidance for people who write and install skills as the first work; those are proposals, not yet published.
One thing it puts in view: the standard allows six frontmatter fields, and a skill with any other field fails to upload to claude.ai or the Skills API. Nine skills in our catalog still have one, shadcn/ui and Impeccable among them; Expo removed it from its three skills on October 28.
NEW & VETTED / 02
REA: point your agent at a binary and ask what it does.
REA (Reverse Engineer Anything) by morluto takes apart shipped software: native binaries, Electron and JavaScript apps, packaged Python, firmware and browser targets. Your agent decompiles, inspects packages and compares versions from the build itself, not from source it may not have. Point it at two releases of an app, for example, to see what actually changed. A-rank · Security · MIT.
Its npm package went from 480 downloads in all of September to 500,000+ in the week ending November 1, with 357 contributors and new releases through October. That growth shows up in downloads and contributors, not just in stars, which is the difference from New Term Radar in 04.
Install it as a plugin: the plugin bundles the skill and registers REA’s own tool server. The skill on its own adds instructions only, not the tools, so we list the plugin. Native analysis uses a Hopper, Ghidra, IDA or Kuna install you already have; JavaScript needs none.
It’s dual-use, and the project says so: its disclaimer limits it to security research, malware analysis, CTFs and software you own or are authorized to examine, and rules out DRM, license-check and anti-cheat circumvention.
THE PICK / 03
Planning with Files — Othman Adi
A-RANK · 27,000+ STARS · 16 CONTRIBUTORS · UPDATED OCTOBER 30, 2026 BY OTHMAN ADI
The fix for losing your place in a long session is to not keep your place in the session. Planning with Files writes the plan to disk: task_plan.md, findings.md and progress.md, in your project. Its first instruction is to read those files back before doing anything else.
What made it the pick this week is how it uses hooks. Before each prompt, before a compaction and when the agent tries to stop, a hook script feeds the current plan back in. So when 01’s cutoff hits, the plan isn’t in the part that gets cut. Version 3.25.2 (October 24; the latest is 3.27.0) also refuses to accept a plan the agent wrote without recognizable phases, a check its completion gate depends on.
The caveats: its own SKILL.md is about 41,000 characters, twice what survives a compaction, so it leans on those files and hooks rather than on its own text. Its frontmatter pre-approves Read, Write, Edit, Bash, Glob and Grep, so in the turn you invoke it, the agent can run shell commands without asking you. Read the hooks in scripts/skill-hook.sh before you install it.
$ /plugin marketplace add OthmanAdi/planning-with-files $ /plugin install planning-with-files@planning-with-files
DIDN’T PASS / 04
Three candidates we checked and turned down.
- Chroma’s MCP serverno longer maintained. It’s Chroma’s own server, but its last commit was September 17, 2025 and its last release August 2025. We don’t add a candidate that’s gone 90 days without a commit.
- XCrawl Skillsinactive, and no license. Submitted through our form. No commits since March 20, and the repository has no license, so it isn’t open source.
- New Term Radaron hold. Also submitted through our form. Its skills need a paid AIsa key and send you to the author’s referral link to get one. The plugin does label it as a referral link, and the skills ask before every paid run. But it went from launch to 2,048 stars and 625 forks in three weeks, and stars and forks are its only sign of growth. They’re easy to inflate, so we’ll look again when there’s evidence beyond them, such as installs or outside contributors.
WORKFLOW / 05
Full Development Lifecycle
Claude connected to your error tracker, issue tracker, GitHub and a memory server, so the context for a task lives outside the conversation. Advanced, about 30 minutes to set up.
Planning with Files keeps one task’s state in local files. This keeps a whole project’s state in four MCP servers: Sentry and Linear for what’s broken and what’s next, GitHub for the code, and Memory for a knowledge graph the next session can read. Each config on the page is copied from the server’s own docs.
WHAT CHANGED / 06
68 skills you may have installed were updated since Issue 008.
“Updated” means the skill’s own directory got a commit since our check on October 19, not a judgment about what the commit did. Before you run update, read the commit.
Fourteen are Microsoft’s Azure and Foundry skills, last changed together on November 2. Nine are Corey Haines’ marketing skills, from Copywriting to SEO Audit. Three are Superpowers (see 01) and three are Expo’s. Top of the rest by stars:
…and 30 more. Each skill’s page shows when it was last updated.
That’s it for Issue 009. See you next week.
— The Manifest