<- all tokdocs

Microsoft SkillOpt-Sleep for Claude Code: What "Zero to Perfect" Means

By , founder of Mavric and Apso

View on TikTok ->

SkillOpt-Sleep is an open-source Claude Code plugin from Microsoft Research that reviews your past sessions overnight, drafts edits to CLAUDE.md and skill files, and keeps only edits that improve a held-out test score. A TikTok from @dripthecup says Microsoft's own test took a weak skill "from zero to perfect in one night" on Claude Haiku. That result is real, but "perfect" means three of three held-out tasks passed a rule check for a risks section and a confidence level, on Claude Haiku 4.5, which Anthropic now lists as a legacy model.

What is SkillOpt?

microsoft/SkillOpt is MIT-licensed and was created on 2026-05-08. On 2026-10-19 it had 18,548 GitHub stars, so the video's figure of 18,328 was accurate for the week it was posted.

The research behind it is the paper "SkillOpt: Executive Strategy for Self-Evolving Agent Skills" (arXiv 2605.23904), submitted 22 May 2026. Microsoft Research published a blog post on it on June 30, 2026.

The video blends two things. SkillOpt the paper trains a skill document against a benchmark using a separate optimizer model. The plugin the video installs is SkillOpt-Sleep, which the repo calls the "deployment-time companion". Its documentation carries a "Preview" notice saying "interfaces and defaults may change", and its latest packaged release on PyPI is 0.2.0 from 2026-07-02.

The plugin README credits three sources for the overnight idea: SkillOpt's validation-gated edits, Anthropic's "Claude Dreams" feature for offline memory consolidation, and the agent-sleep idea. The video presents the framing as Microsoft's alone.

How to install SkillOpt-Sleep in Claude Code

The commands in the video match the plugin README:

git clone https://github.com/microsoft/SkillOpt.git
/plugin marketplace add ./plugins/claude-code
/plugin install skillopt-sleep@skillopt-sleep
/skillopt-sleep dry-run
/skillopt-sleep run
/skillopt-sleep adopt --legacy

The README adds details the video skips:

  • You need Python 3.10 or newer, plus an authenticated claude executable on PATH for the real backend.
  • /plugin install opens a panel where you finish the install with a click, per Claude Code's marketplace docs.
  • /skillopt-sleep status confirms the install.
  • The narration says "type adopt", but the command is adopt --legacy. Runs that edit several skills need adopt --skill NAME or adopt --all-skills.

Clone the repo rather than running pip install skillopt. The README says handoff mode, multi-skill fan-out and reviewed subset adoption "require a source checkout from main until the next release". Every adopt backs up the prior file first.

What did "zero to perfect in one night" measure?

The result comes from the plugin README's results table: backend Claude Haiku 4.5, a seed skill called "brief-writer", held-out score 0.00 to 1.00 in one night. The Codex backend reached the same score in two nights.

The brief-writer seed was written with "no risks section / no confidence level", and the scorer is "a local rule judge". The reproduce command uses --limit-holdout 3. So the jump from 0.00 to 1.00 means three held-out tasks went from missing the required sections to including them. The benchmark is Garry Tan's gbrain-evals skillopt-v1, which uses four deliberately flawed seed skills, all of which went from 0.00 to 1.00.

The model version matters. Anthropic's models overview lists Claude Haiku 5.5 as the current Haiku and files Claude Haiku 4.5 under "Legacy models (still available)", with retirement "not sooner than October 15, 2026". The reproduce command passes --model haiku, so it is not clear which Haiku a rerun today would use.

Does SkillOpt-Sleep improve every skill?

No. Microsoft's own RESULTS.md says the method helps on "recurring tasks with a checkable correctness signal and real headroom" and is "flat within run-to-run noise" on tasks a strong model already handles well. Its main experiment is about safety: GPT-5.4-nano on SearchQA fell from 0.554 to 0.026 over five nights with the validation gate off, and held at 0.570 with the gate on.

The creator's own run matches those limits. On Claude Haiku with the plugin's built-in practice tasks, the score stayed at 0.625 over two nights, with 16 edits drafted and 0 kept. The creator's links page also reports a three-night run at 0.861 that did not change. These are self-reported numbers. The video's card lists an i7-10700F and RX 9060 XT, but with --backend claude the model runs on Anthropic's servers, so the local hardware does not affect the score.

What it costs and what data it sends

The default backend is mock, which the README describes as "deterministic, no API spend". For real learning you switch to --backend claude or --backend codex, which runs "on your own budget". A third option, /skillopt-sleep-handoff run, has the current Claude Code session answer the model calls with no API key.

The video's privacy tip is correct. A real backend "sends truncated transcript excerpts and derived tasks to the selected provider", and "outbound prompts are not currently guaranteed to be secret-free." The docs add three more cautions:

  • Each night writes a local evidence.jsonl with copies of prompts and replies, which the docs say to "treat as sensitive local data".
  • The per-night budget max_tokens_per_night defaults to 400,000 and "is not a hard cap". A real-backend dry-run still costs money.
  • /skillopt-sleep schedule installs a cron job, and adding --auto-adopt lets it rewrite CLAUDE.md without review.

Bottom line

SkillOpt-Sleep is a real Microsoft Research project, and the install steps in the video work. The headline result is three held-out tasks on a section-heading check, run on a legacy Haiku model. Expect little change on tasks your model already does well, and keep --auto-adopt off unless you want unattended edits.

Sources

Published October 14, 2026. Based on a TikTok by @dripthecup. Full video notes (markdown)