Turn a vague idea into a production-grade agent spec — a workflow you can delegate, verify, own, and share.
Specsmith is not a prompt generator. It's a spec-engineering harness: a set of composable Claude Code skills that interview you, classify what kind of agent system your work actually needs, build its context / constraints / evals, adversarially stress-test it, and emit a portable workflow bundle you keep on disk.
It exists because the leverage in 2026 isn't clever prompting — it's the four cumulative disciplines underneath it:
Prompt Craft → Context Engineering → Intent Engineering → Specification Engineering
(table stakes) (curate the environment) (encode what to want) (a blueprint an agent runs solo)
These are cumulative — failures compound downward. A brilliant prompt on a broken context layer still fails. Specsmith diagnoses your weakest rung and fixes it first.
It's built so its structure embodies the principles instead of citing them:
- It right-sizes.
orientdiagnoses you, then routes to a Quick / Standard / Deep subset of skills. A one-shot task doesn't get dragged through a thirteen-phase gauntlet. - It grounds before it advises. Every skill opens by loading the relevant knowledge bundle first (build the data room before the work), not by tacking citations on at the end.
- It produces a habit, not a string. The output is a
workflows/<slug>/bundle — context + spec + constraints + evals + the interaction pattern + a review gate — that you re-run next quarter on a better model and share with a friend. - Evals are first-class. "It finished" ≠ "you trust it." A mandatory judge step proves every stress-test finding became an enforceable requirement.
git clone <your-fork>/specsmith
# the knowledge base ships bundled at knowledge/kb/ — nothing to link.
# (optional) point at your own wiki instead: export SPECSMITH_KB="/path/to/your/wiki"Then, in Claude Code, just describe what you want:
"Help me turn my client-reporting process into something I can delegate to an agent."
The orient skill takes it from there. (Or install as a plugin — see Sharing below.)
| Skill | Job | Discipline |
|---|---|---|
orient |
diagnose the user, route to the right subset | planner |
classify-architecture |
pick the Agent Species + run the complexity-floor gate | architecture |
engineer-intent |
find the real objective + the correctness contract | Intent |
build-context |
assemble the information environment (data room first) | Context |
engineer-spec |
self-contained spec + the meaning layer | Specification |
author-constraints |
MUST/MUST-NOT/escalation + species-scoped enforcement | Specification |
decompose-tasks |
task DAG with a reliability budget (deep) | Specification |
design-evals |
acceptance criteria + A/B/C tests + golden set | Evaluation |
red-team |
hostile review, failure modes, stress tests | Evaluation |
simulate |
dry-run the species across timelines (deep) | Reflection |
optimize |
score prompt variants against the golden set (deep) | Reflection |
audit-feedback-loop |
judge: every finding → an enforceable requirement | judge |
capture-habit |
write the reusable, shareable workflow habit | Reuse |
Specsmith is grounded in a knowledge base distilled from Nate B. Jones's AI-education work (49 concepts, 270 source-video pages). It ships bundled at knowledge/kb/ — source-grounded out of the box, no setup. Point $SPECSMITH_KB at your own wiki to override it, or run on the eight-axiom fallback in knowledge/principles-core.md if you strip it out. Only the curated wiki is included; the raw transcripts are not. See knowledge/KB-LINK.md.
Specsmith owns the upstream (what to build + how to know it's right). Superpowers owns the downstream (build it well). Specsmith hands off to brainstorming, writing-plans, test-driven-development, and verification-before-completion at the seams. See CLAUDE.md.
It's a plain repo of skills — clone it, or install it as a Claude Code plugin (.claude-plugin/plugin.json is included). No personal data, no platform lock-in, no Windows paths. The bundled knowledge base comes with it, so friends get the full source-grounded system out of the box.
Specsmith's substance comes from the public work of Nate B. Jones, an AI strategist and educator who publishes daily, practitioner-focused analysis of where AI is actually useful. The four-discipline ladder, the Agent Species framework (from his Tobi Lütke video, sources/YpPcDHc3e9U), and every insight in the bundled knowledge base are distilled from his videos — each knowledge/kb/sources/<id>.md page carries the original video ID so you can go watch the source. The distillation, the skill architecture, and any errors are the maintainer's; the ideas are his. If this is useful to you, support him directly:
- YouTube — @NateBJones (AI News & Strategy Daily)
- Substack — natesnewsletter.substack.com
- Site — natebjones.com
See docs/ for the frameworks. Built to compose with Jesse Vincent's Superpowers.
MIT — see LICENSE.