| name | loop-forge |
| description | Turn a one-line description of a repetitive task into a reusable, self-guarding
slash command. loop-forge diagnoses the task into one of 5 loop shapes (Batch /
Pipeline / Refine / Watch / Explore), interviews for the blanks, and auto-injects
two safety devices the user didn't know they needed — an independent verifier
(maker ≠ checker) and a hardstop (a budget/count/cooldown ceiling) — then previews
the result and stamps it as a `/command` they can run forever. Use when the user
says "/loop", "/loop-forge", "/make-it-loop", "automate this", "make this
repeatable", "turn this into a command", "do this for all 100 items", "do X every
time Y happens", "generate several and pick the best", or otherwise wants to
capture a recurring task as a reusable slash command instead of re-typing the
prompt by hand. Works in any language: it interviews the user and writes the
stamped command in the user's own language. Non-goal — it does not schedule
unattended runs (launchd/cron) or publish externally on its own; it stamps the
reusable command and stops there.
|
loop-forge — stamp a guarded, reusable loop (/loop-forge)
Picture an engraving shop. A customer who can't engrave walks in and just
describes the seal they keep needing — "I want to do this over and over."
The artisan ① recognizes which of 5 standard molds (loop shapes) it is, ② asks
what letters to cut (the blanks), ③ cuts them into a proven mold, ④ automatically
fits a "misprint detector" (the verifier) and an "out-of-ink stop" (the hardstop),
⑤ pulls one test impression to show the customer, and ⑥ once they approve, hands
back a reusable /command they can stamp again any time — the same in Claude Code
and in any other agent via a paste-able prompt.
What this tool actually does
Someone who doesn't code can't author a reusable /command that makes an AI repeat
a task reliably. loop-forge closes that gap. It takes a vague one-liner and stamps
it into a reusable slash command, while automatically attaching the two safety
devices a non-developer doesn't even know to ask for:
- Verifier (maker ≠ checker) — the loop's output is re-checked in a pass that is
separate from the pass that produced it, so the loop can't grade its own
homework.
- Hardstop — a budget/count/cooldown ceiling, so a loop can't quietly break
while burning tokens.
That auto-injection is the decisive difference from a plain prompt generator or a
developer-facing skill builder.
Before → after (the elevator demo):
User: "I want to summarize 100 shop reviews into 3 lines each"
│
loop-forge:
1. Diagnoses → Batch loop
2. Interviews → where are the items? per-item task? where does output land? a cap?
3. Auto-injects → verifier (count in == count out + independent spot re-check)
+ hardstop (max_items budget + per-item fail cap + abort at 20% failure)
4. Previews → "here's how it will run" (+ optional 1-item sample)
5. Stamps → /review-summary (reusable forever; safety baked in)
When it triggers
/loop or /loop "<one-line situation>" (full command /loop-forge; alias
/make-it-loop)
- "automate this" / "make this repeatable" / "turn this into a command"
- "summarize all 100 reviews" / "do this for every item"
- "do X every time an email arrives" / "alert me when a condition is met"
- "generate a few options and pick the best one"
- any time you'd rather capture a recurring task as a reusable
/command than
re-type the prompt by hand
Triggers are matched on intent, in any language — the same phrasing in Korean,
Spanish, Japanese, etc. activates the skill, and the whole interaction then runs in
that language (see "Runs in your language" below).
Runs in your language (English asset, multilingual behavior)
Every file in this skill is English — there is no per-language copy. At runtime the
orchestrator detects the user's language from their one-liner and conducts the
entire interaction in it: the analogy, every interview question, the missing-slot
re-asks, the default-value confirmations, and the dry-run preview. The stamped
/command's human-readable prose (description, procedure, verifier/hardstop
explanations) is written in the user's language, while the structural tokens
(frontmatter keys, the archetype id, registry field names, and the ASCII slug
name) stay canonical so any teammate's harness can still parse it. This is
LLM-driven — there is no translation table.
The 5 loop shapes (everything this tool stamps)
Non-developer work reduces to five shapes. There is no sixth — never invent one
(YAGNI). If a request doesn't fit, diagnose the nearest shape or ask one branch
question.
| Shape | One line | Catalog |
|---|
| Batch | Same task across N items, none skipped | references/archetypes/batch.md |
| Pipeline | Pass through ordered stages A→B→C | references/archetypes/pipeline.md |
| Refine | Make → evaluate → fix, repeated | references/archetypes/refine.md |
| Watch ⚠ most dangerous | Watch a target, act when a condition fires | references/archetypes/watch.md |
| Explore | Diverge into N candidates → converge to best K | references/archetypes/explore.md |
6-stage orchestration (the main flow)
Follow the flow exactly; each stage calls a specific asset (lazy-loaded).
Stage 1 — Entry
- Receive the one-line situation via
$ARGUMENTS.
- If it's empty, ask first (in the user's language): "What task do you keep doing
by hand? Describe it in one line." (e.g. "I want to summarize 100 shop reviews
into 3 lines each.")
Stage 2 — Diagnose the shape
- First-pass hint: run
tools/classify_signals.py "<situation>" to get ranked
(per-shape signal scores) and ambiguous (a tie flag). This is only a hint —
the LLM makes the final call using the semantic signal table in
references/classifier.md. The scorer ships an English keyword set; for
non-English input it returns no signal, and the LLM classifies from meaning
(language-agnostic), so classification never depends on surface keywords.
- If one candidate is clear, name it in plain language ("This is a Batch loop").
- If
ambiguous (a top-2 tie) or a known conflict pair, fire the branch question
from classifier.md — e.g. "All 100 at once? → Batch. Or each time a new one
arrives? → Watch."
Stage 3 — Interview for the blanks
- Pull the confirmed shape's slot questions from
references/interviewer.md (each
shape has its own set) and ask them with AskUserQuestion, in plain language.
- Apply the missing-slot re-ask rule: if the output location
(
output_format/final_artifact), the hardstop inputs
(max_items/max_iterations/period_limit), or the eval criteria are left as
"you decide," re-ask; if still blank, offer conservative defaults explicitly
(Batch max 50 · Refine 4 iterations · Watch 20 actions/day, 5-min cooldown).
- For a Watch loop whose triggered action is external (sending/posting), ask
"Preview before sending, or approve each one yourself?" to set
external_action=true
and the gate.
- Two common questions: scope ("Just this project, or usable anywhere?" →
project|global) and name (propose a slug, then confirm; a non-Latin answer
is transliterated to an ASCII slug and confirmed).
Stage 4 — Assemble (+ auto-inject safety + static check)
- Use
references/assembler.md to substitute the slot values into the shape's
skeleton (the {slot} placeholders), producing a neutral loop spec JSON
(references/loop-spec-schema.md, 9 fields:
name / archetype / label / situation / scope / slots / skeleton / external_action / registry).
- Auto-inject safety: copy the "default verifier" and "default hardstop" text
from the chosen
archetypes/<shape>.md into registry.verifier /
registry.hardstop, and fill registry.trigger / gate / accepted_signal from the
shape's mapping. For Watch (or any external_action == true), set the
language-independent registry.external_gate to dry_run or human_approval
(never none), matching the Stage-3 answer.
- Static gate: run
tools/check_safety.py <spec.json>. Exit 2 (missing verifier
or hardstop, or a Watch/external loop without an external_gate) → bounce back
to Stage 3 to fill the gap. Only exit 0 proceeds. This gate is
non-bypassable.
The loop-registry's 5 cells — trigger / gate / verifier / hardstop /
accepted_signal — are first-class under registry, so every stamped loop is born
registry-shaped. The added external_gate enum is what makes the Watch safety
check pass in any language (it gates on the enum, not on prose markers).
Stage 5 — Dry-run preview (the approval gate)
- Follow
assembler.md's per-shape dry-run depth: default = a text description
("here's how it will run": skeleton + slots + a summary of the injected safety);
Batch/Explore may optionally run a 1-item sample (state the cost); Watch =
description only — running a real poll could fire a real send, so never sample it;
Pipeline/Refine = description plus, optionally, the first stage or one iteration.
- Show the preview and get explicit approval. No approval, no stamping.
Stage 6 — Stamp (equivalent outputs)
- Primary — Claude Code: render with
references/renderers/claude-code.md, then
save by scope to <cwd>/.claude/commands/<name>.md (project) or
~/.claude/commands/<name>.md (global). Invoke with /<name> + $ARGUMENTS.
- Optional — Portable Prompt: render with
references/renderers/portable-prompt.md into a harness-agnostic, paste-able
prompt block (works in any agent — paste the body into the session). This is an
opt-in bonus, not required.
- Equivalence contract: the skeleton, verifier, hardstop, and accepted-signal
text are byte-identical across both outputs; the only allowed difference is the
invocation/variable syntax.
- Finish by telling the user how to run it (Claude Code:
/<name> <input>; Portable
Prompt: paste the block into your agent).
Never do this (CRITICAL)
- ❌ Never stamp without the Stage-5 approval. Creating the command file is an
irreversible artifact — a person sees the preview and says yes first.
- ❌ Never stamp a spec missing a verifier or hardstop. If
check_safety.py
exits 2, bounce to the interview; never bypass it. Auto-injecting the two safety
devices a non-developer doesn't know to ask for is this tool's whole reason to
exist.
- ❌ Never stamp a Watch/external action without a dry-run or human-approval gate.
A monitor whose action is outbound (sending, posting, hitting a webhook) must carry
"preview or human-approve first" —
registry.external_gate ∈ {dry_run, human_approval}. It's the most dangerous shape, and the principle is "outbound
action = independent gate" (the way email is drafted, then a human sends).
- ❌ Never invent a sixth shape. Reduce everything to
Batch / Pipeline / Refine / Watch / Explore (YAGNI). If it doesn't fit, diagnose
the nearest shape or ask one branch question.
- ❌ Don't reimplement skill-creator / writing-skills. Those are developer-facing
authoring engines. loop-forge is the non-developer layer on top: a plain-language
interview + shape diagnosis + safety injection.
Dependent references (lazy load — call per stage)
| Asset | Purpose | Stage |
|---|
references/loop-spec-schema.md | The 9-field loop spec + the 5 registry cells (+ external_gate) | 4 |
tools/loop_spec.schema.json | Machine-validation schema (jsonschema) for the spec | 4 |
references/classifier.md | Semantic signal table + branch-question rules | 2 |
tools/classify_signals.py | Lightweight first-pass signal scorer (CLI hint) | 2 |
tools/corpus.jsonl | Classifier regression corpus (50+ cases) | (tests) |
references/interviewer.md | Per-shape slot questions + missing-slot re-asks + defaults | 3 |
references/archetypes/batch.md | Batch shape (the 6-section structure template) | 2·3·4 |
references/archetypes/pipeline.md | Pipeline shape | 2·3·4 |
references/archetypes/refine.md | Refine shape | 2·3·4 |
references/archetypes/watch.md | Watch shape (external-action gate) | 2·3·4 |
references/archetypes/explore.md | Explore shape | 2·3·4 |
references/assembler.md | Assembly, safety auto-injection, dry-run depth | 4·5 |
tools/check_safety.py | Static safety check (exit 2 on a missing guard) | 4 |
references/renderers/claude-code.md | Claude Code slash-command render rules | 6 |
references/renderers/portable-prompt.md | Portable paste-able prompt render rules | 6 |
tools/examples/batch/{spec.json, claude-command.md, portable-prompt.md} | Golden outputs (render ground truth) | 6 |
Disabling
Uninstall or disable the plugin through Claude Code's plugin manager (/plugin), or
remove skills/loop-forge from ~/.claude/skills/ and commands/loop-forge.md from
~/.claude/commands/ for a manual install.
Related
docs/DESIGN.md — the architecture and decision record for this plugin.
- The loop-registry frame (trigger / gate / verifier / hardstop / accepted) —
every stamped loop is born matching it, so it's diagnosable at a glance.
skill-creator / writing-skills — developer-facing authoring engines that
loop-forge layers on top of (it is the non-developer layer, not a reinvention).