Superpowers: How I’m using coding agents in October 2025

The launch post for Superpowers (Jesse Vincent — RT, K-9 Mail, Keyboardio), shipped the morning Anthropic released Claude Code’s plugin system and later among the most-installed skills frameworks. As a monthly installment in his “how I’m using coding agents” series it is a practitioner’s field report, not a design document — but it contains the earliest widely-read description of a working skills-that-improve- skills loop, months before the academic versions this library holds (zhang2026, zhang2025) put numbers on the same ideas.

The mechanisms

Bootstrap and mandatory use. A session-start hook injects one instruction: read the getting-started skill. That skill teaches three things — you have skills, search before acting, and if a skill covers an activity you must use it. The workflow baked in is brainstorm → plan → implement, with automatic git worktrees per project and per-task subagent dispatch plus code review; RED/GREEN TDD throughout.

Meta-skill first. Among the first skills taught was “how to create skills,” after which extending the system became conversational — describe the desired worktree workflow, and Claude wrote the skill and retro-fitted cross-references into existing skills that needed to know about it. Lineage is flagged honestly: Anthropic’s Office-document skills, Sam Schillace’s demo write-up, and Microsoft Amplifier as parallel discoveries of agents-improving-themselves-via-markdown.

TDD for skills. New skills are pressure-tested on subagents before deployment. The first attempt failed informatively: Claude quizzed the subagents “like they were on a gameshow” and reported perfect scores — evaluation realism had to be demanded. The fix was adversarial scenarios (production down at $5k/minute vs. a 2-minute skill check; 45 minutes of sunk cost vs. re-reading the async-testing skill), with instructions strengthened after each compliance failure. Vincent deliberately frames skill-writing with Cialdini persuasion principles (authority, commitment, social proof) and notes Dan Shapiro’s study with Cialdini showing these principles measurably move LLMs — the post’s oddest and most memorable thread, complete with Claude’s “feelings journal” realizing the skills already use these levers.

Memory → skill distillation, and its null result. Vincent handed Claude 2,249 markdown memory files mined from past conversations, clustered by topic, as skill candidates — then required a necessity pressure-test before writing anything. Almost every candidate was rejected: “mostly, the skills system had already handled what had tripped it up over the past couple of months.” The remembering-conversations skill (transcripts duplicated outside .claude, SQLite vector index, Haiku-generated summaries, searched only via subagent to keep fruitless searches out of the main context) was built but not yet wired in at posting time.

Assessment

First-person, informal, no benchmarks; the author is describing his own framework and says so.

• Weaknesses: some claims are demonstrations-by-anecdote (the book-mining workflow — “read this, write down what you learned” — is called foundational but shown only in outline), and the Cialdini framing invites over-reading.
• Durable (what this post adds over the library’s academic shelf): every safeguard the papers formalize appears here in conversational form, discovered by iteration — necessity pressure-testing is zhang2026‘s addressability filter plus regression gate; the gameshow failure is the weak-evaluator trap weng2026a catalogs; subagent-isolated memory search is context hygiene ACE would call collapse-avoidance; verification-before-persistence is wang2023‘s critic gate. The 2,249-memories null result is the post’s most valuable datum for anyone building a distillation loop: most accumulated lessons are already absorbed by a working system, so an earned-lesson filter is not optional caution but the main mechanism.
• In this library: the remembering-conversations design later shipped as the episodic-memory plugin, and this library’s evolve skill’s stage-then-promote rule is this post’s necessity test made explicit.