NoBrainer Tech Flow

From task to done.

NoBrainer.tech · Flow overview · Production-ready agentic workflows

Tell it what you need. Flow clarifies the goal, does the work, and checks the result.

Say “Use nb-ultra” (or invoke $nobrainer-ultra in Codex) to fix code, prepare
an everyday document, or investigate a problem. Clear tasks go straight to execution;
meaningful ambiguity gets one focused question round. Done means the agreed
criteria are met and the result is checked. A real blocker is reported with the
next unblock action.

A quick answer stays a quick answer. The model-neutral workflow uses one plan,
bounded corrective attempts and a Markdown checkpoint for longer work. Native
goals, telemetry, subagents and client-specific tools are optional.

Existing installations keep the nobrainer-tech-skills package/plugin ID and all
skill commands. See the Flow migration guide.

Source-backed comparison decisions
cover the current collected Trending snapshot and explain what we adopted or retained.

Try one real task

Install safely, start a fresh session, then give the agent a
small task with a checkable result:

Use nb-ultra. Fix one bug in this project. Reproduce it first, make the
smallest correction and run the relevant check. Tell me what changed and
what remains unverified. Ask only if a missing decision changes the result.

Prefer a non-coding trial? Try an invitation, a small code fix or a bounded
command. The guide gives explicit acceptance criteria so you
can judge your own result. These are trials, not promised benchmark scores.

Install safely

Clone a reviewed ref, validate it, preview exact targets, then apply:

(
  set -u
  : "${NB_REVIEWED_COMMIT:?set a reviewed full 40-character commit SHA}"
  test "${#NB_REVIEWED_COMMIT}" -eq 40 || exit 2
  case "$NB_REVIEWED_COMMIT" in *[!0-9a-f]*) exit 2 ;; esac
  git clone --no-checkout https://github.com/nobrainer-tech/nobrainer-tech-flow.git || exit 3
  cd nobrainer-tech-flow || exit 3
  git checkout --detach "$NB_REVIEWED_COMMIT" || exit 3
  test "$(git rev-parse HEAD)" = "$NB_REVIEWED_COMMIT" || exit 3
  python3 scripts/validate_skills.py --suite || exit 4
  python3 scripts/install_skills.py --client codex || exit 4
  python3 scripts/install_skills.py --client codex --apply || exit 4
)

Set NB_REVIEWED_COMMIT to the exact full commit SHA you reviewed. Tags and
branches are rejected because they can move; every failed gate stops before the
next command.

The installer defaults to all fifteen canonical skills, supports an exact
subset, refuses foreign targets and can use links or copies. Restart the client
and perform clean-session discovery before claiming runtime installation. Full
client-specific steps and rollback are in Installation.

GitHub flow chart

flowchart TD
    A[One outcome] --> B{Material ambiguity?}
    B -->|yes| C[BUDDY: one focused question round]
    B -->|no| D{Small and clear?}
    C --> D
    D -->|yes| E[Direct answer or edit; check the result]
    D -->|no| F[SCOPE + PLAN: outcome, authority, proof; concise TODO]
    F --> G[AUTOPILOT: execute the authorized scope]
    G --> H[Verify; independent REVIEW when useful]
    H -->|verified defect; attempt budget remains| G
    H -->|acceptance met| I[Audit delegated artifacts and stop owned workers]
    H -->|blocked| J[Checkpoint; report one unblock action]
    E --> K[Deliver evidence and stop]
    I --> K
    F -. optional .-> L[Markdown goal for resume]
    G -. independent work .-> M[Bounded native subagents]
    M --> H
    L -. unavailable native goals or telemetry .-> G

Astra Ready Flow, portable by design

Version 1.6 removes client-specific prerequisites from ordinary work.
OpenAI introduced GPT-6 Astra on
September 3, 2026. The suite keeps the host-selected model and uses the same
plain-text instructions with other models; it does not pin a provider or choose
an expensive tier automatically.

Compatibility separates available models, tested
behavior, client loading and source distribution. Local smoke evidence covers
only its recorded source and scenarios. It is not a universal compatibility,
quality or token-savings benchmark.

The 1.6.1 instruction review
clarifies user authority, skill-caused pauses and when verification is sufficient.
The same rules apply across models; historical runtime proof stays tied to the
tested release bytes.

Small task       -> direct result + relevant check
Larger task      -> clarify if needed + short TODO + execute + verify
Resumable task   -> same workflow + one Markdown goal/checkpoint
Independent work -> optional bounded subagents, audited before integration

Load specialists only when needed. Reuse the project's instructions, tests,
specs and wiki. Choose SDD for durable contracts and TDD when a failing test
would expose the behavior; neither requires installing a framework.
See the v1.6 review and research decisions.

Start with one skill

Use nobrainer-ultra for setup or any non-trivial
outcome. A small reversible edit without a public contract, routing, workflow or
portfolio change can use its quick path; other changes use the full path and its
coherence gate.
In Codex, explicit invocation is $nobrainer-ultra; nb-ultra is a
natural-language trigger phrase and therefore depends on a client's implicit
description matching.

DRIFT_CHECK -> BUDDY -> SCOPE -> AUTOPILOT -> VERIFY -> RECEIVE_AUDIT -> LEARN

• BUDDY is the first and only ordinary clarification stage.
• SCOPE freezes outcome, non-goals, expected files, proof, untouched work,
  minimum solution, test decision and clean completion before a non-trivial write.
• One canonical plan owns TODO state; the owner sees only a compact Progress
  checklist and one next action.
• A durable ledger adds exact identity, dependencies, checkpoints, retries and
  rollback only for cross-session, dependency-rich or consequential work.
• AUTOPILOT continues through routine approved work without repeated check-ins.
• nobrainer-team selects the minimum useful capabilities;
  nobrainer-dispatcher schedules only approved ready work in bounded batches;
  nobrainer-sessions creates or reuses exact visible sessions only when
  parallelism, isolation, handoff or independent evidence earns the cost.
• Failed review returns to nobrainer-build; changed code invalidates old proof.
• Merge, deploy, publishing, spending, credentials, destructive operations and
  production mutation remain owner gates unless exact authority is already
  recorded.

The two execution axes stay independent:

CONTROL_MODE: BUDDY -> AUTOPILOT
SESSION_MODE: MAIN | MULTI_SESSION

For resumable or delegated work, SESSION_HEALTH_GATE uses available,
configured limits. Missing optional telemetry lowers the health claim and still
allows bounded safe work. An explicitly required hard budget must be enforceable.
RUNTIME_RELEASE concerns actual task-owned workers; a completed result does not
prove they stopped. GOAL_FILE is optional Markdown recovery state. A native goal
is used only when available and authorized; the file alone suffices.

Autopilot works in one MAIN session. Native subagents need a scoped assignment,
observable completion and reviewed output. Use persistent sessions and a
Dispatcher only when a real handoff or dependent queue needs them.

Fifteen skills, distinct ownership

Aliases are trigger phrases, not duplicate directories. Each skill owns one
recurring boundary:

The curation audit records why each skill exists and
what belongs in another skill instead of becoming a sixteenth trigger.

Correct once, improve permanently

Ultra contains portable semantic hooks for four events:

• OWNER_DECISION_CHANGED updates the canonical decision, marks the old value
  superseded and invalidates dependent TODO items and evidence;
• AGENT_ERROR_CORRECTED fixes the active result and classifies one minimal
  prevention candidate: AUTO_SCOPED may persist it to one governed canonical
  project-local store, ASK prepares one exact diff, and OFF keeps it
  task-local without a durable diff;
• REVIEW_FAILED creates a bounded Build correction and sends fresh evidence
  back to Review;
• REPEATED_DEFECT stops blind retries and invokes RCA with the prior failure
  fingerprint.

Only durable, sourced, authorized and non-secret knowledge is promoted through
nobrainer-wiki. One mutable fact has one canonical owner; the system does not
append contradictory copies to the plan, instructions and wiki.
Project setup records LEARNING_WRITE_POLICY: AUTO_SCOPED | ASK | OFF, so an
owner can enable automatic project-local learning without granting global or
publishing authority.

Quality without AI slop

The shared delivery contract operationalizes:

• KISS: the simplest complete design wins;
• YAGNI: no speculative extension points or future options;
• DRY: deduplicate owned knowledge and state, not incidental similarity;
• SOLID: cohesive responsibilities and stable boundaries without class or
  interface ceremony;
• evidence: no invented APIs, fake runtime claims, placeholder logic, swallowed
  errors, generic prose, broad unrelated rewrites or mock-only confidence;
• content quality: purpose, audience, correctness sources, completeness,
  coherence and target-workflow usefulness are frozen before execution.

If a decision-relevant fact may be current, niche, uncertain, high-stakes or
source-attributed, Ultra routes the smallest sufficient check through Research.
A stable local syntax, import, test or configuration error starts from local
evidence instead of an automatic wiki/web detour. If required primary evidence
is unavailable, it says RESEARCH_BLOCKED instead of guessing.

Compatibility is a proof ladder

All clients consume the same skills/ tree. Thin adapters cover Claude Code,
Codex, Cursor, OpenCode, Gemini CLI, Kimi Code and Pi; the portable Agent Plugin
manifest and project instructions are the fallback for other Agent Skills
consumers.

Compatibility claims use five distinct levels:

SOURCE_VALIDATED -> REPOSITORY_CHECKED -> CLIENT_LOADED
                 -> RUNTIME_VERIFIED -> DISTRIBUTED

A valid manifest does not prove clean-session routing. A local test does not
prove production. See Compatibility for current proof
and Testing for acceptance evidence.

Current source version: 1.7.1. Check the
latest published GitHub release
for distribution and the v1.7.1 rebrand record for the
preserved installation identities. The unchanged command runner keeps its
v1.7.0 verification scope. Source publication does not
imply client marketplace discovery or improved model reasoning. The earlier
v1.6.1 publication readback
remains historical evidence.

Version v1.5.0 remains an
accepted rollback source release. Its historical pre-publication checkpoint is
docs/releases/v1.5.0.md.

Version v1.4.0 remains the
previous accepted source release and rollback option. Client-specific runtime
rows remain evidence-scoped; source publication does not imply marketplace
discovery.

The v1.3.1 release remains the
previous accepted source release and a rollback option. It adds the English
BRIEF writing mode, concise issue/story templates and surface-specific bug
evidence.

The v1.3.0 release remains an
older accepted source release and rollback option.

Version v1.2.1 remains the rollback source release
at full commit 0010140d19a7ff847dff776569772ef04d82c314, with its own exact
tag, archive and isolated-install evidence. To reproduce the reviewed v1.3.0
source:

git checkout --detach 8ae4a26548ce908fc5f98b22663f52e163541f56
test "$(git rev-parse HEAD)" = "8ae4a26548ce908fc5f98b22663f52e163541f56"
python3 scripts/validate_skills.py --suite
python3 -m unittest discover -s tests -q

v1.2.0 remains published but failed archive acceptance; its exact boundary is
recorded separately. GitHub reports the v1.2.1
release object as non-immutable and tag protection was not independently
verified, so security-sensitive consumers should pin the full commit SHA. The
current tag archive was re-read after the metadata-only history rewrite and
passed 88/88 tests; historical CI binds the same tree, not the current commit
identity.
v1.1.0 remains the accepted rollback anchor at full
commit 711be31d654835a04ef8c70674c3e493aeb2da8a.

One source, thin adapters

skills/               canonical portable behavior
adapters/bootstrap.md small session-start route to Ultra
hooks/                tested client lifecycle adapters
scripts/              validation and conflict-safe installation
tests/                deterministic and behavior-contract gates
docs/                 compatibility, curation, eval and release evidence
assets/               brand and workflow diagrams

Client-specific forks of a skill are prohibited. External skills discovered by
Team are untrusted input: inspect the exact source/ref, instructions, scripts,
permissions, network/credential behavior, trigger overlap and rollback. Prefer
temporary project-scoped use; persistent or global installation is an owner
gate.

Attribution

• nobrainer-autoimprove is an independent adaptation of Andrej Karpathy's
  autoresearch measure-change-keep
  loop.
• nobrainer-wiki is an independent adaptation of Andrej Karpathy's
  LLM wiki concept.
• Other external design sources remain cited inside the exact skill where they
  influenced a behavior contract.

Contributing and security

Read CONTRIBUTING.md, SECURITY.md and
RELEASE-NOTES.md. Every change goes through a focused PR,
pressure scenario, validators, diff review and secret scan. Do not report
publication, distribution, live routing or user-visible success without
readback from that layer.

NoBrainer Tech builds practical agentic workflows for teams that want speed
without surrendering control. Learn more at nobrainer.tech
or browse ready-to-use workflow products on
Gumroad.