"Iterative improvement loop — refine a target across multiple Algorithm cycles toward ideal state
Most AI improvement workflows are single-pass: generate, review, done. That works when the target is simple and 'better' is obvious. For anything that requires genuine iteration — a skill that's almost right but not quite, a prompt that needs tightening across multiple dimensions, a document that compounds quality with each pass — one cycle isn't enough. But running the same thing again manually loses the learning from the prior cycle. There's no memory of what was tried, what was rejected, and why.
Loop runs the Algorithm in mode: loop — multiple full OBSERVE→THINK→PLAN→BUILD→EXECUTE→VERIFY→LEARN cycles on the same target, with each cycle's LEARN phase feeding directly into the next OBSERVE. The ISA tracks iteration count, cumulative improvements, and a Dead Ends section so rejected approaches don't get retried. Human review happens between cycles by default, so you can redirect before the next pass. --autoresearch mode drops the human gate for overnight runs: it borrows the pi-autoresearch pattern of no-review-between-cycles plus MAD confidence gating, which flags marginal iterations (delta < 1.0× MAD of iteration scores) as noise-floor rather than letting them update the baseline. --iterations sets the maximum cycle count; without a clear exit condition or score threshold, the loop runs to that cap.
What you say to your DA, and what the Loop skill actually does.
The thinking, frameworks, and architecture that distinguish this skill from a generic version of the same task.
/loop runs the Algorithm in mode: loop — multiple full Algorithm cycles on the same target, each iteration building on the last. By default a human reviews and redirects between iterations. Unlike /optimize (an autonomous mutation loop), /loop runs full Algorithm passes with that human review in the seam.
Some work doesn't finish in one pass. A skill, a prompt, a diagram, a piece of writing gets meaningfully better each time you run a full cycle on it — but only if each cycle remembers what the last one learned and what it already tried. Run the cycles by hand and you lose that thread: you re-explore dead ends, forget which approaches got rejected, and have no record of whether the score actually moved. /loop carries ISC criteria and a dead-ends ledger across iterations so each pass starts from where the last one ended.
Each iteration is a full Algorithm cycle (OBSERVE → LEARN). The LEARN phase of one cycle feeds the OBSERVE phase of the next, the ISA tracks iteration count and cumulative improvements, and a human approves or redirects between iterations unless autoresearch mode is enabled.
/loop --target "path/to/target" --iterations 5
/loop --target "~/.claude/skills/Art/Workflows/TechnicalDiagrams.md" --goal "make diagrams more consistent"
/loop --resume # Resume a previous loop
/loop --status # Show iteration history
Each iteration is a full Algorithm cycle (OBSERVE → THINK → PLAN → BUILD → EXECUTE → VERIFY → LEARN) with:
| Argument | Required | Default | Description |
|---|---|---|---|
--target PATH |
yes | What to improve (file, directory, skill) | |
--goal TEXT |
inferred | What "better" means for this target | |
--iterations N |
3 | Maximum number of Algorithm cycles | |
--resume |
Resume a previous loop | ||
--status |
Show iteration history | ||
--autoresearch |
off | Opt-in autonomous mode — see below |
Sets mode: loop in ISA frontmatter. The iteration field tracks cycle count. Each cycle re-enters the Algorithm with accumulated context from prior iterations.
--autoresearch switches /loop from supervised multi-pass improvement to autonomous iteration, borrowing three patterns from pi-autoresearch (davebcn87, MIT):
--iterations reached, target met, or explicit interrupt.## Dead Ends section. Every failed iteration appends one line with the rejected approach and reason. Resumes read this to avoid retrying rejected paths.|delta|/MAD(iteration_scores) per cycle. Flag red (<1.0×) iterations as noise-floor and log marginal; do not update baseline. See PAI/ALGORITHM/optimize-loop.md → Confidence Gating.Invocation:
/loop --target "path" --goal "X" --iterations 20 --autoresearch
Default /loop behavior is unchanged — autoresearch is opt-in only. Intended for overnight runs on targets where human-in-the-loop review between cycles is too slow.
/loop --target "~/.claude/skills/Research" --goal "improve output quality" --iterations 5
/loop --target "prompts/summarize.md" --goal "more concise, less filler"
Say any of these to your DA and PAI activates the Loop skill automatically:
Or invoke explicitly:
Skill("Loop")The thinkers, books, frameworks, and research this skill is built on. The ideas belong to them — the integration belongs to PAI.
Install PAI on your machine — your DA gets the Loop skill plus 44 others, all hooked into one Life OS.