nobrainer-tech-skills

Portable agentic workflows for fast, visible and evidence-gated delivery.

NoBrainer.tech · NoBrainer workflow blueprints and field guides

One canonical set of Agent Skills for Claude Code, Codex, Cursor, OpenCode and
GitHub Copilot. The clients get thin adapters; the operational truth stays in
skills/.

Continuous improvement beats delayed perfection. Start with the smallest
safe workflow, verify the result, keep only durable learning, and improve the
measured bottleneck. A clear daily task stays in one session; specs, a wiki,
extra sessions and evaluation loops appear only when they earn their cost.

NoBrainer is the control plane around delivery: it decides what should happen,
when work is ready, which session owns it, what evidence is required, and when
the workflow must stop. Official Superpowers remains the external implementation
methodology for planning, worktrees, TDD, debugging, review, and verification.

The Karpathy-inspired learning loop stays explicit and inspectable:

• nobrainer-wiki compounds sourced project knowledge, decisions and confirmed
  user preferences across sessions;
• nobrainer-autoimprove turns repeated corrections or measured gaps into a
  bounded baseline/eval/change/keep-or-revert experiment;
• nobrainer-ultra retrieves only relevant knowledge and adapts the workflow,
  without silently rewriting instructions or building an opaque user profile.

Start here

Use nobrainer-ultra (nb-ultra) for a non-trivial
task. It chooses the smallest workflow that can deliver the outcome:

DRIFT_CHECK -> BUDDY -> READY_GATE -> AUTOPILOT -> VERIFY -> RECEIVE_AUDIT

• BUDDY is a short requirements/decision gate, not a permanent chat mode.
• AUTOPILOT executes the approved scope without asking between routine steps.
• A small task stays small: no SDD, wiki, worker or improvement loop by default.
• Visible sessions are added only when handoff, isolation, resume or real
  parallel benefit outweighs coordination cost.
• Merge, deploy, publishing, spending, credentials, destructive actions and
  production mutation remain owner gates.

Nine skills, no filler

Aliases are trigger phrases in each description; they are not duplicate skill
directories. The installer adds the complete curated set because every active
skill has a distinct role in normal delivery.

The rationale and retirement rules for every active skill are recorded in the
skill curation audit.

Compatibility and proof

Portable source, adapter wiring, clean-session routing, and marketplace
distribution are separate proof levels. See Compatibility
for the acceptance transcript required to promote any client to runtime-verified,
and Installation for client-specific setup and readback.

Why one repository

Client-specific repositories would duplicate the protocols and drift. This repo
uses the same approach as a well-structured plugin bundle:

skills/              canonical portable skills
.claude-plugin/      Claude packaging
.codex-plugin/       Codex packaging
.cursor-plugin/      Cursor packaging
.opencode/           OpenCode registration adapter
.github/             Copilot repository instructions

Retired predecessors exist only in Git history, so plugin discovery and the
installer expose only the nine canonical skills.

Superpowers: complementary, not copied

NoBrainer Tech Skills owns lifecycle, visible sessions, owner gates,
specification contracts, decision/RCA records, measurable improvement and wiki
knowledge. Official Superpowers owns
implementation methods such as brainstorming, planning, worktrees, TDD,
systematic debugging, requesting/receiving implementation review and
verification. nobrainer-review owns the final evidence-gated finding contract,
not another implementation framework.

Install Superpowers separately for each client. This repository deliberately
does not vendor or rename its skills, preventing duplicate triggers and stale
private wrappers.

Specialist discovery without permanent bloat

NoBrainer Ultra first checks this complete curated set and the task's native
tool. Only a real capability gap may route to the open
skills CLI:

npx skills find "$SKILL_QUERY"
npx skills use "$SKILL_SOURCE@$SKILL_NAME"

skills use is preferred for one-off evaluation because it does not make a
permanent installation. External skills are untrusted input: inspect the exact
source/ref, SKILL.md, scripts, license, write scope and secret/network behavior
before use. Persistent or global installation remains an explicit owner gate.
There is deliberately no separate NoBrainer “find skills” wrapper; discovery is
a fallback inside nobrainer-ultra, not another always-installed trigger.

The retired monolithic autopilot, fixed ten-agent RCA, duplicated wiki helpers,
custom multi-review harnesses, client/account integrations, broad security
catalog, 401-agent team builder, old Ultracode and old Karpathy wrappers are
absent from the active tree and remain recoverable from Git history.

Installation

The commands install the exact Git ref you checked out. Before applying writes,
confirm that checkout contains skills/nobrainer-ultra/SKILL.md and
scripts/validate_skills.py.

Clone the repository, validate it and dry-run the portable installer:

git clone https://github.com/nobrainer-tech/nobrainer-tech-skills.git
cd nobrainer-tech-skills
python3 scripts/validate_skills.py --suite
python3 scripts/install_skills.py --client codex
python3 scripts/install_skills.py --client codex --apply

The installer writes all nine curated skills, and nothing else, unless repeated
--skill NAME flags request an exact subset. It writes nothing until --apply,
refuses to overwrite existing targets and supports Claude, Codex, OpenCode,
Copilot and the shared
~/.agents/skills convention. Cursor uses the plugin manifest. Full commands,
Superpowers setup, dynamic discovery and rollback are in
docs/INSTALL.md.

Quality gates

The core contracts have frozen pressure scenarios and independent review in
docs/evals/core-suite-2026-08-27.md,
plus a baseline/candidate setup comparison in
docs/evals/setup-upgrade-2026-08-27.md.
These are local contract checks, not production or buyer-outcome proof.
Deterministic repository checks:

python3 scripts/validate_skills.py
python3 scripts/validate_skills.py --suite
python3 -m unittest discover -s tests -v

The validator checks frontmatter, directory/name identity, aliases, public-clean
content, relative links and retired-name exclusion. Behavioral tests remain
necessary: valid Markdown does not prove a workflow makes the right decision.
See Testing for the four evidence layers and CI boundary.

Design sources and attribution

• nobrainer-autoimprove is an independent adaptation of Andrej Karpathy's
  autoresearch experiment loop.
• nobrainer-wiki is an independent adaptation of Andrej Karpathy's
  LLM wiki note.
• Superpowers is created by Jesse Vincent / Prime Radiant and remains an
  external dependency under its own project and license.

Contributing

Read CONTRIBUTING.md and AGENTS.md. Change one
behavioral concern at a time, start with a failing behavior/eval, preserve
public-clean boundaries, run the full local validation, and use a focused branch
and PR. CLAUDE.md must remain byte-identical to AGENTS.md. Report security
issues through SECURITY.md, not a public issue.

License

MIT © 2026 NoBrainer.tech