← NXT Step home

The STACK, proven

Everyone is building AI agents, skill packs and second brains. We studied about 75 of them: courses, open-source repos and paid programs. We kept the parts that held up on real work and measured the result.

~2,340tokens saved on every message by cutting the main instruction file from 352 to 163 lines
224 → 36skills cut, after finding most had never been used
13 → 0leaked keys in settings once a script started blocking them
3/5 → 5/5answers the right size in a blind test once a hook checked them

One bet behind the whole system

Every piece of our system has to earn its place: each skill, memory, rule and agent. Code checks that on every run, so the system stays small enough to stay sharp.

S

Source

Memory and the brain
How most setups do itSave everything. Chat history, a notes app, or every email and message piped into one searchable pile.Seen in second-brain workshops, memory courses and firehose builds.
What goes wrongThe pile drowns the signal. Old notes outvote new ones, and every chat starts slower than the last.
How ours worksOne fact per file. A small set of always‑on rules, a router that loads only what the job needs, and memories that fade unless they get used.
ProofHot memory index cut from 25.9 KB to 13.9 KB while the vault grew.Routing test: all 9 sample jobs open the right file.
T

Tools

Skills and connectors
How most setups do itInstall a skill pack. 25, 33 or 47 skills at once, plus whatever the course hands out.Seen in open-source skill packs and paid skill bundles.
What goes wrongSkills fight each other. Each one adds to every prompt, the wrong one fires, and nobody knows if any of them help.
How ours worksOne entry skill per discipline with sub-skills that cost nothing until called. Each skill is tested on the same job with and without it, then re-tested every month.
ProofTook 6 of 47, 1 of 33 and 0 of 25 from three popular packs.Copy skill: 14/14 checks passed with it, 12/14 without.
A

Approvals

Gates and guardrails
How most setups do itWrite the rules into the prompt: "never do X", "always check Y". Then hope.Seen in most courses and starter templates.
What goes wrongA rule in a prompt is a request. The same mistake comes back the week the AI is busy with something else.
How ours worksRules that matter become scripts that run on every action and block the mistake. Nothing is called "live" until a check proves it.
ProofLeaked keys 13 → 0. A repeat deploy failure 7 → 0.A gate people skipped 58 times was deleted, too.
C

Cadence

Learning and routines
How most setups do itFix the output in the moment. Maybe keep a list of gripes to add to the prompt later.Seen in skill courses and self-improving system videos.
What goes wrongThe fix lives in one chat. Next week the AI makes the same mistake, and the gripe list never gets used.
How ours worksCorrections are caught at the end of every session and saved as rules. A rule broken twice becomes a gate. Scheduled jobs leave a receipt as proof they ran.
ProofFirst review of 166 past sessions found 9 fixes the live capture had missed.14 gaps closed in the copy checks after reviews found them.
K

Keys

Agents and ownership
How most setups do itBuild a team of agents with an org chart, often on a hosted platform that holds the keys.Seen in agent workforce programs and hosted agent apps.
What goes wrongThe agent was allowed to do it. Most agent failures trace back to what it was allowed to do.
How ours worksAn agent is AI with a job, your facts and rules it can't break. Each one has one lane and says it is an AI. Anything that sends, spends or deletes passes a gate we own. Reviewers grade its work against real outcome data.
ProofOur audit found 87 of 90 agents never used, so we cut to the ones doing real work.Safety gate: 18/18 test cases.

A system that gets sharper as it grows

Mistakes come in as rules and graduate to gates. Unused skills and stale memories leave on a schedule. That loop is what keeps it sharp.

Where we learned it

Every source below was reviewed against our own system. Credit where it's due.

Took and improved
  • Austin Marchese: self-improving loops, which became our session heartbeat and turn gate.
  • Jake Van Clief: a short router file that points to the right instructions, now tested with sample jobs.
  • Eliot Prince: testing skills with and without, plus real example pairs.
  • Karpathy's wiki pattern: notes compiled into cited pages.
  • superpowers, SkillSpector, qmd and stop-slop: kept, then tuned to our rules.
Took the idea
  • Anthropic's four loops: showed we only had timed loops, so we built one that runs every turn.
  • Boris Cherny's archetypes: a role is a mode one agent switches into.
  • Agent workforce programs: approvals and ownership are the missing layers.
  • Paid AI-brain webinars: about 12 skills is the right size for a pack.
  • Nick Saraev: his auto-ingest second brain made the case for keeping memory small and curated.
Turned down, and why
  • Big skill packs: mostly duplicates that crowd every prompt.
  • Tools that squeeze command output: they hide the evidence our checks need.
  • Hosted agent apps with unclear privacy terms: we keep the keys.
  • Cloud-only brains: one sync error can delete a folder.