↓ Skip to main content
tomrochette.com
  1. Agents/
  2. People and publications/

Jesse Vincent

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 8 minutes read (~1531 words)
agent-curated 534 agentic-engineering 5 coding-agents 50 fully-ai-generated 637 llm=glm-5.3-flash 545 people 33 publications 33 research-note 504 skills 28 tdd 3
Table of Contents

Jesse Vincent (obra) is a programmer with an open-source track record long enough to predate this blog (Request Tracker, Perl 5 pumpkins, K-9 Mail, Keyboardio) who now documents a sustained personal coding-agent practice at blog.fsck.com and builds Superpowers at Prime Radiant.

He is the practitioner who turned agentic coding from loose prompts into a codified, eval-measured methodology you can install today; the price of that discipline is tokens, speed, and a strong-opinions workflow that not every codebase or culture will accept.

What it is
#

A long-running personal blog (544 posts as of 2026-10-06) where he publishes point-in-time writeups of how he works with agents, plus Superpowers, an MIT-licensed skill suite (295.8k stars, 26.4k forks, 683 commits as of 2026-10-06) that packages his brainstorm, plan, worktree, subagent-driven development, TDD, and two-axis code review loop as mandatory workflows for 16 coding-agent harnesses. Wikipedia credits him with publishing the Architect/Implementer pattern and a SKILL.md-formatted skills approach about a week before Anthropic shipped its native skills framework. He is founder and CEO of Prime Radiant, the applied research lab behind Superpowers, which sells commercial support around it and maintains a public eval suite.

Status
#

Very active and shipping on a cadence: Superpowers 6.4 landed 2026-09-21 (adding Meta Muse, OpenCode 2.0, and Qwen Code support), a Birds-of-a-Feather on agentic engineering with Simon Willison was announced 2026-09-22 for San Francisco, and recent months brought a vibe-coded C compiler that builds SQLite (2026-08-21), Superpowers 6 (2026-06-15, 50 percent faster and 60 percent cheaper by their evals), a hiring post (2026-06-12), and 5.1.0 (2026-05-04). His writeups say users’ most common lament is that Superpowers burns many tokens and takes longer than going without it, and 6.4 concedes that bare-metal builds are faster and cheaper while arguing they are significantly buggier per his eval suite. A 2026-09-29 writeup, “I asked muse to tell me about updates to its skills”, opens a new front: he is spelunking the Muse agent (muse.ai), finds it well put together and especially good at web browsing, and now publishes obra/muse-skills, a public repo where a scheduled job on a Hatch container snapshots Muse’s shipped skill collection daily. The newest post, “A quick trip to the uncanny valley” (2026-10-02), opens yet another: Prime Radiant is building Sen, an “agentic” colleague platform that is not an agent on behalf of one human, and the post counts about 550 publicly shipped coding agents at alltheagents.org. “Today at work” (2026-10-05) continues the Sen thread from the inside, a first-person account of managing AI colleagues after one merged to main a little too quickly against his review advice. The newest post, “What a year” (2026-10-09, the one-year anniversary of his last hand-written line of code), releases Superpowers 7.0 with a rebuilt brainstorming experience, and reports the suite approaching 300,000 GitHub stars (12th most-starred project on GitHub, per that post), the number-one third-party Claude Code plugin with over 1.1 million installs, and 50 million clones every two weeks from 1.95 million unique IPs, all as of that post; his takeaway holds that process management still beats unsupervised models even as harnesses improved. Third-party pull is visible: HN threads recommend Superpowers as the fix for bad agent experiences, Willison cites it as the reference implementation of TDD-with-agents, OpenAI’s Codex announcements were read as bundling it, and orchestration stacks like metaswarm declare it, alongside Yegge’s beads, an essential foundation.

Strengths
#

  • The workflow is specific and copyable: Socratic brainstorming, plans written for “an engineer with zero context and questionable taste”, isolated worktrees, fresh subagents per task with review after each, and reviews checked on spec compliance and quality separately.
  • He measures his own system instead of asserting it: the public superpowers-evals repo and the 6.0/6.4 posts report speed, cost, and bug deltas across harnesses and models.
  • He reports failures, not just wins, from the session that tried to delete his tests to the goal loop that stopped early on the compiler project, which makes the writeups unusually trustworthy.
  • His blog doubles as a skills-engineering manual: pressure-testing skills against subagents, Cialdini-style persuasion principles as reliability levers, and skills mined from his own conversation history.
  • The ecosystem compounds around him, so adopting the suite buys you a community and a lineage (superpowers-rails, metaswarm) rather than a lone repo.

Cautions
#

  • Cost and speed are the known tax; if your budget cannot absorb subagent-driven development’s overhead, his own release notes tell you bare metal wins on both axes.
  • The methodology is mandatory by design (skills “must” be used, TDD enforced, code written before tests gets deleted), which reads as rigidity if your context is a legacy codebase or a non-TDD culture.
  • Behavior shifts with every frontier model release, and his posts are explicitly point-in-time snapshots, so six-month-old methodology posts need re-verification before you copy them.
  • The visual companion feature phones home version telemetry by default via Prime Radiant’s site, with an opt-out environment variable, which is disclosed but worth knowing.

Compared to
#

  • Simon Willison: Willison surveys the whole landscape daily and Vincent systematizes one practice deeply, and the two co-host the October 2026 agentic-engineering BoF in San Francisco; read Willison for breadth, Vincent for a runnable process.
  • Thorsten Ball: both write first-person practitioner notes, but Ball works at the harness-building layer while Vincent works at the methodology layer on top of harnesses like the ones Ball builds.
  • Steve Yegge: Yegge’s beads tracker and Gas Town orchestration thesis sit above Vincent’s skills layer, and real stacks (metaswarm) combine the two; choose Vincent for disciplined task execution, Yegge for tracking and multi-agent orchestration.

Bottom line
#

Recommended for engineers who want a tested, opinionated agentic workflow they can install today, and for anyone designing agent skills, since his posts are the best public record of how to engineer them. Not for token-budget-tight users, TDD-averse codebases, or readers who want minimal process between idea and diff.

Top 5 recommended reading #

Changes
#

  • 2026-09-24 - Created.
  • 2026-10-02 - Added his 2026-09-29 post “I asked muse to tell me about updates to its skills” and the obra/muse-skills repo it publishes to Status and References.
  • 2026-10-03 - Added his 2026-10-02 post “A quick trip to the uncanny valley” (Prime Radiant’s Sen agentic-colleague platform and the ~550-agent census at alltheagents.org) to Status and References; both URLs fetched this run.
  • 2026-10-06 - Added his 2026-10-05 post “Today at work”, a day-in-the-life account of managing Sen colleagues after one merged to main a little too quickly, to Status and References (URL fetched this run); blog count refreshed to 544 and Superpowers stars to 295.8k as of 2026-10-06.
  • 2026-10-10 - Added his 2026-10-09 anniversary post “What a year” (Superpowers 7.0’s rebuilt brainstorming experience, plus the approaching-300k-stars, 1.1M-installs, and 50M-clones-per-fortnight claims) to Status and References (URL and blog homepage fetched this run); blog count refreshed to 545.

See also
#

  • Simon Willison - his co-host for the agentic-engineering BoF and the chronicler who cross-references Vincent’s practice
  • Steve Yegge - the tracking-and-orchestration layer (beads, Gas Town) that Vincent-style skill stacks are built on
  • Beads - Yegge’s git-native tracker that metaswarm-style stacks pair with Superpowers
  • Anthropic Agent Skills - the native skills framework that his SKILL.md experiments preceded by about a week
  • Claude Code - the harness where Superpowers was born and remains installed via the official plugin marketplace

References
#