OPEN SOURCE FIRMWAREGITHUB TEMPLATE · MIT · VIEW SOURCE →
fullbuild.ai HARNESS

HARNESS FIRMWARE

BOOTSWARM

Give every repo its own memory system
Harness stores project-specific rules, lessons, and workflows in the repo, so new sessions start informed.

CREATE A PROJECT → VIEW TEMPLATE project knowledge survives session resets
REFERENCES · repo facts read when needed →
FIG 01 · SESSION STATE DECAYS · REPO MEMORY REMAINS
MEMORY project lessons survive the chat VERSIONED WITH EACH REPO 23 workflows ready when needed ON DEMAND · SOURCE LINKED 2 coding agent runtimes CLAUDE CODE · CODEX MIT open source and inspectable GITHUB TEMPLATE
NEW PROJECT · WEB · WINDOWS · MACOS

Create the repo and the system together

Name the project once. Harness creates the repo with memory, workflows, and both agent setups already inside.

CREATE A PROJECT → GET THE LOCAL LAUNCHER
REPO MEMORY SKILLS ADAPTERS
STEP 02 · ASSEMBLE Harness arrives with the repo

Project memory, 23 workflows, Claude Code setup, and Codex adapters are already there.

PROJECT READY RUN INIT-PROJECT
STEP 03 · BUILD Open it and start working

Run init-project once to tailor the system to the repo.

WEB: AUTHORIZE GITHUB ONCE · LOCAL: REUSES YOUR EXISTING GH LOGIN · PRIVATE REPO BY DEFAULT

COMMITTED PROJECT KNOWLEDGE · ILLUSTRATIVE MEMORY FLOW · pitfalls.md · 2026-07-18

Stop solving the same repo problem twice

Save a lesson once. /recall brings it back when a later task needs it and shows where it came from.

STEP 01 · OBSERVE A session finds a problem

Stale local server · volatile trace

SAVE LESSON pitfalls.md
STEP 02 · COMMIT Save the fix in the repo

Named reference · reviewable repo change

KNOWN CHECK AVAILABLE
STEP 03 · RETRIEVE Future sessions reuse it

Same repository or clone · retained guidance

MEMORY FLOW · ILLUSTRATIVE

claude · session A · 2026-07-18

preview shows an outdated version of the site

server reuse on a busy port · stale process from a prior session

/recall save local preview servers: stale or wrong site

saved → .claude/reference/pitfalls.md · reviewable repo change

session ended · context cleared

agent · later matching task · same repo or clone

/recall routes to .claude/reference/pitfalls.md

reads cited entry · known check available · port owner checked first

DEFAULT FILE-BASED LAYER · RUNTIME-SPECIFIC

Keep every session lean

A small dense green phosphor core inside a vast faint residue halo; the core shows the 9,355 bytes Claude loads at startup beside 188,216 bytes of skill instructions available on demand

CLAUDE: 5,401 B + 3,954 B = 9,355 B · 9.1 KiB · ≈2,339 tok/turn · CODEX: 3,252 B + 3,954 B = 7,206 B · ≈1,802 tok/turn

Only a small index loads at startup. The 188,216-byte skill library stays out of context until a task needs it.

FOUR OF THE STANDING RULES, VERBATIM FROM CLAUDE.md
  • Reproduce bugs before fixing them.
  • Never claim visual/UI verification you didn't actually perform.
  • Can't run the authoritative check, flag the risk plainly, don't claim it passes.
  • Scope discipline: No unrequested refactors, features, abstractions, or extra coding.

CLAUDE.md adds 5,401 B at startup · read it before adopting

DEFINING OPERATING LOOPS · 8 workflows · 23 in the full library

Eight workflows do the heavy lifting

pan or point · arrow keys select a skill

Twenty-three phosphor bands comparing the byte size of each canonical SKILL.md instruction file; the largest is 26,286 bytes and the smallest is 3,111 bytes

writing-skills · 26,286 B · main instruction file

Dot count shows relative file size. The 23 main skill files total 188,216 B, or 184 KiB. Support files are separate.

They remember fixes, improve the harness, start repos, share updates, cut context, and verify risky work.

  1. recallsave and retrieve repo lessons
  2. refinefix the harness after friction
  3. init-projectset up a repo from detected facts
  4. sync-startershare fixes without overwriting local rules
  5. optimize-contextremove guidance that costs more than it helps
  6. long-horizoncarry one contract across long work
  7. impartial-reviewask fresh agents to check the work
  8. wow-loopimprove visuals through screenshot review

23/23 generated adapters present at Codex tree a137de9f · contract checks cover adapter sync · CI exercises the Windows project generator end to end

CONTEXT ARCHITECTURE · SMALL ROOTS · FOCUSED SUBFILES

Load only what the work needs

CLAUDE.md and AGENTS.md stay small. They point each agent to focused project files only when the work needs them.

HOW WE OPTIMIZE CONTEXT →
SESSION START · ALWAYS LOADED
CLAUDE CODE CLAUDE.md Core rules and routing index CODEX AGENTS.md Runtime-specific boundary
ROUTING LAW Match the task, then open the relevant file
REFERENCES + SKILLS · LOADED WHEN MATCHED
architecture.mdsystem map commands.mdbuild and test deployment.mdrelease facts pitfalls.mdhard-won lessons secrets.mdsafe secret map tech-stack.mdtools and versions SKILL.mdtask workflows

ROOT RULES STAY SHORT · REPO FACTS LIVE IN REFERENCES · WORKFLOW DETAIL LOADS ON DEMAND

WHAT MAKES IT POWERFUL

Built to improve with use

The repo keeps what works. Every improvement stays visible, reviewable, and easy to undo.

  1. Project facts live in the repo Architecture, commands, deployment details, and pitfalls survive every session.
  2. The right lesson comes back /recall finds and cites the relevant file. It saves new lessons in the right place.
  3. Fix the workflow, not just the output /refine turns recurring friction into one small, reversible change.
  4. New repos start from facts /init-project reads the stack, asks only what it cannot infer, and removes irrelevant setup.
  5. Write each skill once Claude stays canonical. Generated Codex adapters add runtime-specific discovery and safety rules.
  6. Share fixes without erasing local rules /sync-starter moves generic improvements while protecting project-specific memory.
  7. Use fresh eyes for hard work long-horizon, impartial-review, and wow-loop add independent checks when the runtime supports them.
RUNTIME SUPPORT · SHARED SYSTEM · EXPLICIT ADAPTERS

Built for Claude Code and Codex

Harness Firmware supports Claude Code and Codex from one canonical skill source Twenty-three canonical workflows pass through a runtime adapter and reach separate Claude Code and Codex setups with different measured resident sizes. CANONICAL SKILL SOURCE WRITE ONCE 23 WORKFLOWS ADAPT PER RUNTIME RUNTIME 01 CLAUDE CODE CANONICAL WORKFLOWS 9,355 B MAX RESIDENT RUNTIME 02 CODEX GENERATED ADAPTERS 7,206 B RESIDENT
FIG 07 · ONE PLAYBOOK SOURCE · RUNTIME-SPECIFIC SETUP

Write each skill once. Harness keeps Claude canonical and generates Codex adapters with runtime-specific rules. All 23 adapters are present. rev d9cd99f5 · canonical tree 503356de · Codex tree a137de9f

FLASH IT · MIT · GitHub template

Start warm

NEW REPO · NEW PROJECT LAUNCHER

launcher: Windows or macOS · project name: my-app

existing GitHub CLI login detected

Creating private repo 'my-app' from template ryanportfolio/Harness-Firmware ...

> gh repo create my-app --template ryanportfolio/Harness-Firmware --private --clone

Stripping template-only files and replacing README ...

> git -C my-app push -q

DONE. Private repo created, cloned, and pushed

next: open it in Claude Code · run /init-project

create from the web or get the local launcher for Windows and macOS · the local launcher reuses your signed-in GitHub CLI session

SKILLS ONLY · EXISTING REPO

/plugin marketplace add ryanportfolio/Harness-Firmware
/plugin install claude-starter@claude-starter

installs 23 skills · 184 KiB across the main skill files

names: Harness Firmware is the template · claude-starter is its plugin on the marketplace · fullbuild.ai is the site you are on

Use the template for the full system. Run /init-project to fill in the repo facts. Use the plugin commands when an existing repo only needs the Claude Code skills.

BEFORE YOU FLASH · OPERATIONAL ANSWERS

Questions the repo answers

How much context does it add?

Up to 9,355 bytes (about 2,339 tokens) for Claude Code and 7,206 bytes (about 1,802 tokens) for Codex. Tool catalogs and runtime-added context are not included.

How does memory persist?

/recall saves lessons in the repo and brings them back when a later task needs them.

Which runtimes use the same source?

Claude holds the canonical skill. A repo script generates the Codex adapter. Runtime-specific rules stay separate.

What does 184 KiB count?

The 23 canonical SKILL.md files total 184 KiB. Their 162,181 B of support files and 31,801 B of generated Codex adapters are separate.

Can I use it in an existing repo?

Yes. The marketplace commands add the Claude Code skills. Use the template when the repo also needs Codex adapters, references, and sync wiring.

What do I need installed?

The web creator only needs GitHub. The local Windows and macOS launcher needs Node and Git. An authenticated GitHub CLI session makes private repo creation automatic.

Does it use my existing GitHub login?

Locally, yes. The launcher reuses your existing GitHub CLI login. On the web, you authorize the Harness GitHub App once, then later projects can be created in one step.

POWER-ON · NEW PROJECT · MIT

Boot the next repo warm

CREATE A PROJECT →