Skip to content
Conduit

skill

techneering

Spec-driven development skills plugin for Claude Code and Codex: 16 tn: skills for spec, implementation, verification, release, and archive.

devnomad-byte/techneeringcommunitylistedunreviewednot running here yet

Set up

Skills are installed by placing their folder where your agent loads skills.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below; one-click install arrives with accounts.

README

synced Oct 14, 2026

⚒️ Techneering

Engineering, with ancient discipline.

Write the spec. Strike while it's hot. Ship with confidence.

license GitHub stars skills platform claude codex

tn: · a skills plugin for Claude Code and Codex

English · 简体中文


The problem

A model coding on its own will drift. It guesses when it should ask. It over-builds when it should stop. It forgets decisions made ten minutes ago, declares success without checking, and leaves you to discover — in review, in prod, at 2 a.m. — what it quietly got wrong.

The output looks like code. It isn't craft.

The discipline

Techneering wraps the model in a harness — Anthropic's word for the scaffolding around an agent — so its effort stays pointed where you aimed it. Sixteen skills walk a change from idea, through spec and code, to a verified, archived release.

The spec is the contract. The harness is the discipline. Every line of code traces back to a requirement; nothing ships on a hunch.


Four layers, one system

AI tooling has evolved through four layers. Most stop at one. Techneering was built for all of them.

LayerQuestionIn Techneering
PromptHow to write one instruction wellevery SKILL.md is a carefully designed prompt
ContextHow to manage what the AI seesthe workspace — files carry state, not chat memory
HarnessHow to constrain what the agent doesthe 16-skill scaffold: spec = contract, harness = discipline
LoopHow to get the AI to iterate to qualityassemble → redgreen → gate → audit → inspect → ship → vault

These aren't separate features — they're the same system seen from four angles. And Techneering keeps the parts loop systems usually skip: explicit failure limits, generation ≠ evaluation, and a depth dial so a trivial tweak skips the full ceremony.


The 16 Skills

The names follow a forge workshop: the spec is raw material, the code is the forged product, and each skill is a station on the shop floor.

🧭 The Smithy — core flow

SkillCallWhat it does
compasstn:compassGateway. Auto-loaded at session start — picks the right path
scouttn:scoutThink out loud. Explore ideas, write no files
drafttn:draftGenerate proposal + specs + design + tasks in one pass
forgetn:forgeImplement. Chains blueprint → assemble internally
audittn:auditDual-layer check: code↔spec + code↔real requirements
vaulttn:vaultMerge delta specs into main specs, archive the change

🧠 The Drawing Board — thinking & planning

SkillCallWhat it does
sparktn:sparkBrainstorm design, detect coding style, recommend frameworks
blueprinttn:blueprintBreak the spec into bite-sized, testable tasks

🔨 The Anvil — execution

SkillCallWhat it does
assembletn:assembleDispatch one subagent per task, two-stage review each
redgreentn:redgreenTest-driven: red (failing test) → green (minimal code) → refactor
diagnosetn:diagnoseSystematic debugging: root cause before any fix

🔍 The Inspection Bench — quality gates

SkillCallWhat it does
gatetn:gateIron law: run the command, read the output, then claim it works
inspecttn:inspectGlobal cross-task review — consistency, architecture, security

📦 The Shipping Dock — utilities

SkillCallWhat it does
isolatetn:isolateCreate an isolated git worktree, verify the test baseline
shiptn:shipBranch integration: merge / PR / keep / discard
crafttn:craftFrontend design guidance (auto-stacked on UI keywords)

Quick Start

Techneering runs on both Claude Code and Codex. Same skills, same tn: prefix — pick your host.

Claude Code:

# 1. Register the marketplace
claude plugin marketplace add https://github.com/devnomad-byte/techneering.git

# 2. Install — project-scoped is recommended
claude plugin install tn@techneering -s project

# 3. Start Claude Code and strike
claude

Codex:

# 1. Register the marketplace (clones the repo as a source)
codex plugin marketplace add devnomad-byte/techneering

# 2. Install
codex plugin install tn@techneering

# 3. Start Codex and strike
codex

Then type /tn:scout to take it for a spin. On every session start, tn:compass is auto-injected (via the SessionStart hook on either host) — just say what you want to build and Techneering routes you to the right station.


How It Works

Techneering picks one of four paths based on change size, then walks you through it station by station:

PathFlow
Large changespark → draft → forge → gate → you choose verify depth
Medium changedraft → forge → gate → you choose verify depth
Urgent bugdiagnose → redgreen → gate → you choose verify depth
Trivial tweakdirect edit → gate (optional)

Always available:

  • scout — pure thinking, writes no files
  • craft — auto-stacked when frontend work is detected

Inside forge, two skills chain automatically — blueprint breaks the work into tasks, assemble dispatches a fresh subagent per task with spec-compliance + code-quality review after each.

OptionFlowWhen
1 Full verificationaudit → inspect → ship → vaultProduction code
2 Standard flowaudit → ship → vaultRoutine features
3 Quick commitship → vaultPrototypes, small changes
4 Verify onlyauditJust check, don't commit

Every skill writes its progress to disk. Close the terminal, come back tomorrow — Techneering scans the files and resumes exactly where you struck off.


Design Philosophy

You drive the AI. Not the other way around.

PrincipleWhat it means
You hold the reinsEvery step is a handoff you approve. AI proposes, you decide. Nothing ships without a human checkpoint.
The spec is the contractRequirements live in files, not chat. Every line of code traces back to a SHALL/MUST — drift becomes visible, not invisible.
Generation ≠ evaluationThe agent that wrote the code never grades it. Separate reviewers (tn:audit, tn:inspect) check against spec and reality — models grade their own work too leniently to self-evaluate.
Context is a file, not memoryEach skill reads its inputs from disk and writes its outputs back. New session, new subagent, same source of truth — nothing lives only in someone's head.
Simplest harness that worksFour paths, not one. Trivial tweaks skip the full ceremony; only production code walks every gate. Complexity is added when the task earns it, never before.

This pattern — decompose, hand off via artifacts, separate generation from evaluation — comes from Anthropic's work on long-running agents. Techneering applies it to human-in-the-loop development: the harness keeps the AI disciplined, you keep the harness pointed at the right target.

The loop is the latest layer — a pattern Addy Osmani named in 2025: stop hand-prompting every step, design the system that does it. Techneering was built this way from the start — including the parts most loop systems skip: explicit failure limits (tn:diagnose stops after 3 failed fixes), generation ≠ evaluation (the writer never grades its own work), and a depth dial so trivial tweaks skip the full ceremony.


The Workspace

Running tn:draft generates a self-contained workspace at your project root. It is the source of truth; nothing lives only in chat.

techneering/
├── config.yaml          # tech stack · coding style · quality rules
├── specs/               # authoritative module definitions
├── changes/
│   └── <name>/          # one directory per change
│       ├── .tn.yaml
│       ├── proposal.md  # why  — the motivation
│       ├── specs/       # what — SHALL/MUST · WHEN/THEN
│       ├── design.md    # how  — decisions & trade-offs
│       └── tasks.md     # todo — checkbox-tracked, resumable
└── archive/             # finished changes, fully preserved

tn:audit traces every line of code back to a requirement. tn:vault merges finished specs into the main specs/, so knowledge compounds instead of evaporating.


Requirements

Pick one host (Techneering works on either, or both):

DependencyWhy
Claude Code CLI or Codex CLI (latest)the host that runs the skills
Git 2.20+the isolate station needs worktrees
Bash 4.0+the session-start hook (shared by both hosts)

Runs on Windows 10/11 · macOS 12+ · Linux.


Star History

↑↓ navigate openesc close