iterate

iterate

Self-correcting agentic loop for Claude Code.

Work → Evaluate → Iterate → Done.

Claude executes your task, evaluates its own output against your success criteria, diagnoses failures, applies targeted fixes, and repeats — until every criterion passes or max iterations are reached.

Why

Research shows most users write vague, underspecified prompts — and benchmarks don't reflect this reality. When you say "make it work" or "looks professional," a single-shot LLM just guesses what you mean. iterate turns that vague intent into a structured quality loop:

1. Infers what "done" means — decomposes vague goals into testable criteria
2. Actually verifies its work — reads files back, runs code, checks assertions (not self-assessment from memory)
3. Fixes what's wrong — targeted corrections, not full restarts

No one has shipped this as a reusable Claude Code skill. This is it.

Install

From GitHub:

claude plugin install jatinmayekar/claude-code-iterate

Local testing:

claude --plugin-dir /path/to/claude-code-iterate

Usage

/enhance — Fix the prompt before execution

/enhance <vague task description>

Takes your vague prompt, identifies what's missing, rewrites it into a structured spec with testable criteria, and asks for confirmation before executing. Best for when you know what you want but not how to ask for it.

/enhance build me a sales dashboard

→ Finds 6 gaps (no metrics, no design spec, no data, ...) → generates 7 testable criteria → asks to confirm → executes → verifies.

/iterate — Self-correct the output after execution

/iterate <task> | <criteria, comma-separated> [| max:<N>]

Examples

Code generation with quality criteria:

/iterate write a Python function that sorts a list | handles empty list, handles duplicates, has type hints, includes docstring

Fix failing tests:

/iterate fix the failing test in test_auth.py | all tests pass, no new warnings

Creative writing with constraints:

/iterate write a haiku about recursion | exactly 5-7-5 syllables, references recursion, no cliches

Custom iteration limit:

/iterate refactor the auth middleware | no breaking changes, passes all tests, under 100 lines | max:8

No criteria (auto-inferred):

/iterate add input validation to the signup form

Claude infers success criteria and asks you to confirm before starting.

How It Works

1. WORK — Execute the task (or apply targeted fixes on iteration 2+)
2. EVALUATE — Check every criterion independently: [PASS] or [FAIL] with evidence
3. DECIDE — All pass? Done. Any fail? Diagnose and fix. Max reached? Report partial.
4. DIAGNOSE — Root-cause each failure, plan the fix, assess risk to passing criteria
5. Repeat until done

Smart Features

• Auto-infer criteria — Skip the | and Claude proposes testable criteria for your approval
• Stuck-loop detection — Same criterion fails 3x with same root cause? Claude asks you for help instead of burning iterations
• Targeted fixes only — After iteration 1, only fixes what's broken. Doesn't redo passing work
• Honest self-evaluation — Won't rubber-stamp its own output

Tested

16 tests across code generation, charts, HTML landing pages, PowerPoint presentations, React components, SVG diagrams, algorithms, and BigCodeBench-Hard benchmark tasks. Tested with Opus, Sonnet, and Haiku (including low-effort mode).

Self-correction in action:

Test	What happened	How iterate helped
SQL Builder	Invalid import on iteration 1	Caught error, fixed on iteration 2
K-Means (BigCodeBench/92)	sklearn unavailable	Reimplemented K-means from scratch using numpy
Orbital Mechanics	Assertion tolerance too tight	Verified physics was correct, diagnosed test spec issue
Contradictory Criteria	return 42 and return 99 both required	Creative solution using file-based state

Bug found during testing: The evaluation step was too weak — Haiku rubber-stamped its own output. Fixed by adding verification rules (read files back, run code, inspect actual output instead of evaluating from memory).

Full results: tests/TEST_RESULTS.md

Output

═══ ITERATE COMPLETE ═══
Status: ALL CRITERIA MET
Iterations: 3 / 5

Criterion Results:
  [PASS] handles empty list — returns [] for empty input
  [PASS] handles duplicates — [3,1,3,2] → [1,2,3,3]
  [PASS] has type hints — def sort_list(items: list[int]) -> list[int]
  [PASS] includes docstring — Google-style docstring with examples

Iteration Log:
  #1: Full implementation → 2/4 passed (missing type hints, docstring)
  #2: Added type hints and docstring → 3/4 passed (docstring lacked examples)
  #3: Added examples to docstring → 4/4 passed