"The Case Against Skills" — @lauraentis
Every / Context Window · 2026-07-16
Why this is in the vault
RDCO runs 20+ Claude Code skills as the backbone of its always-on COO agent. This piece provides the first empirical benchmark (SWE-Skills-Bench) quantifying skill-library bloat risk — 39 of 49 public skills had zero impact, 3 actively degraded performance — and lays out the audit framework for what survives. Directly applicable to RDCO's harness architecture.
The core argument
Skills are reusable packages of instructions (sometimes tools + examples) that load when relevant. The format was popularized by Anthropic and spread on X, where large skill libraries became status symbols. Every's head of tech consulting Mike Taylor argues this has become counterproductive:
Why skills often fail now:
- Frontier models (Fable 5, GPT-5.6) have absorbed the capabilities that older skills were patching. Skills that compensate for a model's blind spot become redundant — or harmful — the moment the next model version closes that gap.
- Adding skill instructions that conflict with how the model was trained ("fighting the weights") increases error rates.
- Cost inflation is real: the worst-performing skill in the benchmark increased token use by 451% with no output improvement.
Empirical backing — SWE-Skills-Bench: Tested 49 public software-engineering skills across frontier agents:
- 39 skills: no impact on performance
- 3 skills: actively made things worse
- 7 skills: improved outcomes — all provided specialized knowledge the model couldn't supply on its own (financial-risk formulas, traffic-management rules)
What survives the audit (the durable category): Skills that give the model information it couldn't have known because it's private:
- Personal preferences and writing style
- Company-specific templates or brand guidelines
- Internal data or exact workflow sequences
- Custom tool access tied to internal systems
What expires: Skills that patch general model weaknesses. These have a shelf life tied to the model version, not the task.
Recommended audit framework (Mike Taylor's approach):
- Keep: private context, custom tool access, personal taste, company-specific workflows
- Retest: skills compensating for model quirks — likely outdated
- Retire: skills that don't demonstrably improve results (test with/without)
- For skills borrowed from others: only trust those that are actively maintained and pruned
Practitioner signal (Skill Share section): Naveen Naidu (Monologue GM) has dropped most generic skills — "because the model got so good." The one public skill that earned its place: OpenClaw's autoreview skill, which packages changed code and sends it to a separate model for review. This is durable because it encodes a workflow (cross-model validation loop), not a general instruction. It enabled a nine-hour overnight build shipping a full feature across backend, Mac, iPhone, and web with zero human code review intervention.
Inside Every: Austin Tedesco (head of growth) has abandoned skills entirely except for "compound engineering" — a plugin providing reusable agent workflows for planning, completing, reviewing, and learning from work. Again: survives because it's a private workflow, not a generic instruction.
Curation section
ONE LAST THING — What's on Every's radar:
- Thinking Machines / Inkling: Mira Murati's post-OpenAI lab dropped its first open-weight model. Positioned as American-made alternative to cost-efficient Chinese open-weight models, not a direct frontier competitor.
- OpenAI companion device: Bloomberg reports a screenless smart speaker serving as a humanlike AI companion — manages home appliances, answers messages, runs on advanced ChatGPT Voice Mode. Still in development.
- Siri on iOS 27 beta: Engineering lead Andrey Galko finds it markedly better at speech recognition and smarter than any local model on iPhone. "Apple will do what they always do — take good technology and make it mass market."
Mapping against Ray Data Co
RDCO's /process-newsletter skill is the clearest test case for the article's audit framework. It passes because it encodes private workflow (RDCO-specific frontmatter templates, vault path conventions, bias-detection logic, sponsor classification rules) — not generic "write well" instructions. That's exactly the category the SWE-Skills-Bench found to be durable.
The broader RDCO harness likely has mixed exposure. Skills like morning-prep, check-board, and process-inbox encode private workflows with RDCO-specific tool integrations (Notion MCP, Gmail MCP, vault paths) — these are safe. Skills that are primarily tone or style guidance (e.g., "write concisely") should go through the audit: test with/without against golden outputs and retire if delta is negligible.
The autoreview pattern (cross-model review loop as a skill) is a direct architectural complement to RDCO's verify-dispatch / verify-vault-write pattern — both use a fresh-eyes second pass as the durable mechanism. Worth formalizing this as a first-class design principle in the harness: skills that encode review loops outperform skills that encode instructions.
The SWE-Skills-Bench finding (only 7 of 49 public skills helped) is a useful prior against adopting community skills wholesale. Any new public skill install should pass the private-context or durable-workflow test before going into the harness.
Related
- [[synthesis-harness-thesis-dissent-2026-04-12]] — Vault dissent note on the "thin harness, fat skills" thesis; this Every piece provides the empirical data the dissent was missing
- [[2026-04-19-indydevdan-cracked-claude-agent-skills]] — IndyDevDan's positive case for skills as building blocks; read alongside this piece for the full architectural tension