Two companion Agent Skills from Saido Labs that turn Claude into a disciplined, self-driving build partner: loop gives every project a repeatable Plan → Build → Test → Reflect → Improve cadence backed by a persistent .md wiki, and goal lets you declare a measurable outcome and have Claude run that cadence until the outcome is verifiably met.

They work in Claude Code (CLI) and Cowork, and follow the open Agent Skills standard so they port to any tool that supports it.

loop-goal-skills/
├── loop/                       # the /loop skill
│   ├── SKILL.md                # cadence, phases, operating principles
│   ├── references/
│   │   ├── phase-playbooks.md   # per-phase checklists, anti-patterns, hand-offs
│   │   └── state-schema.md      # field-by-field spec for LOOP_STATE.md + wiki/
│   ├── assets/
│   │   ├── LOOP_STATE.template.md
│   │   └── wiki/                # starter wiki (index, overview, decisions, …)
│   └── scripts/
│       └── loop_init.py         # scaffolds state files into a project folder
├── goal/                       # the /goal skill
│   └── SKILL.md                # objective-mode entry point that drives the loop
├── loop_goal_skills/           # pip package: `loop-skills` + `loop-init` CLIs
├── pyproject.toml              # packaging (pip install from GitHub)
├── claude-settings.json        # optional pre-approved tool permissions (Claude Code)
├── LICENSE                     # MIT
└── README.md

A reusable operating cadence for any project or build pipeline. Every unit of work moves through five phases — Plan → Build → Test → Reflect → Improve — and each pass writes both working output and a short reflection, so the next pass starts smarter and nothing is lost between sessions. State lives in two places per project:

• LOOP_STATE.md — the cockpit: current iteration, phase, the one active goal, open loops, last reflection.
• wiki/ — durable knowledge: decisions, architecture, learnings, glossary.

loop is a meta-skill: it sequences the work and hands off to specialist skills (docs, slides, engineering, etc.) rather than replacing them.

The objective-mode entry point. You state one or more measurable outcomes, and Claude runs the loop toward them — logging every run, looping back on failure, and stopping only when each target is verifiably met (or a guardrail trips). The whole thing hinges on one rule: every outcome needs a definition of done that TEST can check with evidence.

/goal
1) ParryAI injection classifier ≥95% recall on the held-out injection eval
   — done when recall ≥ 0.95 and precision within 2pp of baseline
2) Document the final infra + architecture in wiki/architecture.md  — after #1
mode: auto
stop: #1 verified twice on held-out data, then do #2; if 6 iterations gain <0.5pp recall, pause and report

Built-in integrity guardrails: no "pass" without citing the real number, verify on held-out data, watch the metric you could regress, and pause on a plateau or a genuine blocker instead of spinning.

Install straight from GitHub, then run the bundled installer to copy both skills into your Claude skills directory:

pip install git+https://github.com/Jcapathy/loop-goal-skills.git

loop-skills install             # → ~/.claude/skills/   (loop + goal, all projects)
loop-skills install --project   # → ./.claude/skills/   (current repo only)

Start a new Claude Code session and invoke with /loop or /goal (they also auto-trigger when your wording matches their description).

The package also installs loop-init, which scaffolds a project's persistent state from the bundled templates:

loop-init ./ParryAI --name ParryAI --mission "Prompt-injection armor for LLM apps"
# creates ParryAI/LOOP_STATE.md and ParryAI/wiki/

No pip? Copy each skill folder into your skills directory — user-level (all projects) or project-level (one repo):

# user-level (available everywhere)
cp -r loop ~/.claude/skills/loop
cp -r goal ~/.claude/skills/goal

# or project-level
mkdir -p .claude/skills && cp -r loop goal .claude/skills/

Start a new session; the skills are auto-discovered. Invoke with /loop or /goal (they also auto-trigger when your wording matches their description).

Each skill folder can be zipped to a .skill package and installed with the Save skill button, or added via Settings → Capabilities. Invoke with /loop / /goal.

Both SKILL.md files declare an allowed-tools line so Claude can use Bash, Read, Write, Edit, Glob, Grep, and friends without a permission prompt while the skill is active. For persistent, always-on pre-approval, merge claude-settings.json into your ~/.claude/settings.json (or a project .claude/settings.json). It fully allows the file/search/web tools and scopes Bash to a common dev stack (python, pip, pytest, node, npm, git), with a deny block for destructive commands and secret files.

⚠️ Merge the permissions block into an existing settings.json — don't overwrite the whole file. Bash command-scoping enforcement has known edge cases in Claude Code; for fully hands-off runs, launch with --permission-mode acceptEdits (or, in a contained repo, --dangerously-skip-permissions).

Autonomous within a run, persistent across runs: the LOOP_STATE.md + wiki/ files let a new session pick up the same objective and run ledger exactly where the last one left off.

Version history is tracked in CHANGELOG.md, following Keep a Changelog and SemVer. Latest: v0.1.0.

Released under the MIT License — © 2026 Saido Labs LLC. Free to use, modify, and redistribute with attribution.