Hermes supports self-rewriting skill files in PM and marketing agent demos

• According to aakashgupta's 100K-stars post and the official Hermes repo, Hermes is pitching itself as a self-improving agent that creates and updates skills from experience, not just a chat session with tools.
• In aakashgupta's six-week log, a recurring competitive briefing dropped from 20 minutes to 8 minutes while the underlying skill file rewrote itself four times.
• shannholmberg's workflow thread and shannholmberg's diagram sketch the same operating pattern: prototype in Hermes, let the harness write and refine the skill, then move the stable version into Claude Code or a VPS.
• For creators, QualitativAi's hackathon demo turned that pattern into an AI filmmaking stack with job packets, browser-native execution, taste memory, and human review at the end of the loop.
• The official v0.12 release notes add one more piece: a background Curator that grades, consolidates, and archives skills on a 7 day cycle.

You can read the official skills docs, check the SOUL.md personality docs, and see that scheduled jobs can attach one or more skills in the cron docs. The interesting part is how quickly users started treating that stack as creative infrastructure, from weekly PM briefings to a browser worker that shuttles image jobs through Flow and back into a searchable taste archive.

The core claim is simple and unusually concrete: Hermes writes to ~/.hermes/skills/ during real runs, then uses the updated file on the next run. The official skills system docs say all bundled, hub-installed, and agent-created skills live in that directory, and that the agent can modify or delete any skill.

That lines up with aakashgupta's statelessness post, which frames the bet as persistent procedure rather than a better base model. The prompt assembly docs also show where the persistence comes from: SOUL.md loads first, followed by frozen MEMORY and USER snapshots, then the skills index.

Across the PM and marketing demos, the workflow has three stages:

1. Prototype the task in Hermes against real work.
2. Let the harness capture corrections, error recovery, and longer tool runs into a skill.
3. Move the stable workflow into Claude Code or a scheduled VPS job.

shannholmberg's model routing post adds another detail worth stealing: different steps route to different models inside the same workflow, with Opus for copy, GPT for planning, GPT Image for visuals, and Codex or Kimi for coding.

QualitativAi's hackathon build is the clearest creator example in the evidence pool. The system starts with a Creative DNA pack and project vault, uses Kimi for lore work, turns approved directions into structured job packets, then hands execution to a bounded browser worker that runs Flow and returns renamed assets with lineage intact.

The back half is where it gets interesting. According to QualitativAi on pre-evaluation, the Brain scores and ranks images before review; according to QualitativAi on taste memory and QualitativAi on lightboxes, accepted and rejected outputs become searchable visual memory; and in QualitativAi's holdout test, the blind holdout score moved from 3 of 31 to 7 of 31 after taste seeding.

The official product surface is catching up to the demos. In the v0.12 Curator release, Nous says Hermes now runs an autonomous background Curator on the gateway's cron ticker, grades the skill library, consolidates overlaps, prunes dead skills, and writes per-run reports.

That matters because the community examples are already hitting skill-library sprawl. petergyang asking about Hermes versus OpenClaw reads like a market check, but the more revealing artifact is the release itself: Hermes is no longer just promising self-rewriting skills, it is adding a maintenance loop for the pile of skills those runs create.