Improve Loop

AutoImprove Loop Skill

Every experiment runs in an isolated git worktree. The main codebase is never modified during experiments. Only winning changes get squash-merged back.

Main branch ──────────────────────────────────── (never touched mid-session)
                 │              │
           experiment-001  experiment-002
           (kept ✅ → merge) (discarded ❌ → deleted)

Pre-flight checks

Before the first iteration, print each check as you run it:

━━━ Pre-flight ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
✓ Config found
✓ Git working tree clean
✓ Base commit: abc1234
✓ Worktree directory ready
✓ Baseline score: XX/100
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

1. Check .claude/autoimprove/config.md exists. If not, stop: "Run /autoimprove:setup first."
2. Check git is available: git status
3. Confirm main working tree is clean. If not, stop: "Please commit or stash changes before running autoimprove."
4. Record the base commit: git rev-parse HEAD — all experiments branch from here.
5. Run the worktree skill's setup step to create .claude/autoimprove/worktrees/ and update .gitignore.
6. Run the measure skill in the main directory to get the BASELINE score.
7. Report: "Baseline: XX/100. All experiments will run in isolated worktrees. Main branch is safe."

Session Header

After pre-flight passes, write a session header to .claude/autoimprove/log.md:

## Session — [ISO 8601 timestamp]
**Planned:** N iterations
**Focus:** "focus string" (or "all improvement areas" if none)
**Baseline:** XX/100
**Base commit:** [full SHA]
**Status:** IN_PROGRESS (0/N completed)

If the log file doesn't exist, create it with the project header first:

# .claude/autoimprove/log.md

> Generated by [autoimprove](https://github.com/benmarte/autoimprove) — Claude Code Plugin
> Project: **[project name]** · Stack: [detected stack] · Started: [date]

---

Then append the session header.

Continue Mode

When invoked with continue-mode parameters (from the /autoimprove:continue command), the loop behavior changes:

• start_iteration — Start numbering from this value instead of 1
• total_iterations — Use this as the display total (e.g., "Iteration 5/10")
• session_mode — If continue, skip creating a new session header; instead update the existing one:
  • Update **Planned:** to the new total if it changed
  • Update **Status:** to IN_PROGRESS
• baseline_score — If provided, skip baseline measurement and use this value
• experiment_offset — Start experiment numbering from this value to avoid branch name collisions

In continue mode, the pre-flight still runs (clean tree, config check, worktree setup) but skips creating a new session header and optionally skips baseline measurement.

Progress Updates

CRITICAL: At the start of every step, you MUST output a visible progress line to the user. Do not silently run tools — always print status first. Use this format:

━━━ Iteration N/TOTAL ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🔬 STEP_NAME: brief description of what's happening

Example progress lines:

━━━ Iteration 1/5 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🔬 PROPOSE: Targeting error handling in src/api/client.ts
🔬 SNAPSHOT: Measuring BEFORE score...
🔬 IMPLEMENT: Adding try/catch to unhandled async calls
🔬 MEASURE: Measuring AFTER score...
🔬 DECIDE: 85 → 89 (+4 pts) — KEPT ✅
🔬 LOG: Recorded to .claude/autoimprove/log.md

Never run more than one step without printing a progress line. The user must always know what iteration you're on and what phase you're in.

The Loop

Step 1 — CREATE WORKTREE

Use the worktree skill to create a new isolated branch and directory:

EXPERIMENT_ID=$(printf "%03d" $N)
git worktree add -b "autoimprove/experiment-$EXPERIMENT_ID" \
  ".claude/autoimprove/worktrees/experiment-$EXPERIMENT_ID"

All work for this iteration happens inside .claude/autoimprove/worktrees/experiment-$EXPERIMENT_ID/. The main directory is not touched.

Step 2 — PROPOSE

If a FOCUS string was provided: Every iteration targets that specific focus. Break the focus into file-by-file or function-by-function sub-tasks and tackle one per iteration. Do not rotate to other areas — stay on the focus until all iterations are used or the focus is fully addressed.

If no FOCUS was provided: Choose one focused improvement from the Improvement Areas in .claude/autoimprove/config.md. Rotate areas — don't repeat an area that failed last time.

State the hypothesis explicitly:

"I will [specific change] in [file(s)] because I expect [metric] to improve by ~[X] points."

Step 3 — SNAPSHOT (BEFORE score)

Measure from inside the worktree directory (same commands, different cwd):

cd .claude/autoimprove/worktrees/experiment-$EXPERIMENT_ID
# run measurement suite from .claude/autoimprove/config.md

Record as BEFORE.

Step 4 — IMPLEMENT

Make the change inside the worktree. The main directory is untouched. Commit the change to the experiment branch:

cd .claude/autoimprove/worktrees/experiment-$EXPERIMENT_ID
git add -A
git commit -m "experiment($EXPERIMENT_ID): $HYPOTHESIS_ONE_LINE"

Step 5 — MEASURE (AFTER score)

Run the full measurement suite again from inside the worktree. Record as AFTER.

Step 6 — DECIDE

If AFTER > BEFORE — KEEP ✅

Squash-merge the experiment back to main:

cd [main project root]
git merge --squash "autoimprove/experiment-$EXPERIMENT_ID"
git commit -m "autoimprove($EXPERIMENT_ID): $HYPOTHESIS_ONE_LINE

Score: $BEFORE → $AFTER (+$DELTA pts)
Files changed: $FILES"

# Clean up
git worktree remove ".claude/autoimprove/worktrees/experiment-$EXPERIMENT_ID"
git branch -D "autoimprove/experiment-$EXPERIMENT_ID"

If AFTER == BEFORE — KEEP ✅ only for clear readability wins, DISCARD otherwise

Same merge process as above if keeping, discard process if not.

If AFTER < BEFORE — DISCARD ❌

Main branch is already untouched. Just delete the worktree:

git worktree remove ".claude/autoimprove/worktrees/experiment-$EXPERIMENT_ID" --force
git branch -D "autoimprove/experiment-$EXPERIMENT_ID"

No rollback needed — there was nothing to roll back.

Step 7 — LOG

Append to .claude/autoimprove/log.md in the main directory:

## Iteration N — [timestamp]
**Hypothesis:** [what you tried and why]
**Branch:** autoimprove/experiment-NNN
**Files changed:** [list]
**Before:** [X/100] — type: X, build: X, tests: X, lint: X
**After:**  [X/100] — type: X, build: X, tests: X, lint: X
**Decision:** KEPT ✅ (merged to main) / DISCARDED ❌ (worktree deleted)
**Reason:** [one sentence]

Step 7b — UPDATE SESSION STATUS

After logging the iteration, update the session header's **Status:** line:

1. Read .claude/autoimprove/log.md
2. Find the last line matching **Status:** IN_PROGRESS
3. Replace it with **Status:** IN_PROGRESS (N/M completed) where N is the current iteration number and M is the total planned
4. Write the file back

If the status line cannot be found, append a warning to the log and continue — the iteration records are the source of truth.

Step 8 — REPEAT from Step 1

Session end — cleanup

After all iterations (or if the user stops early):

Update the session header status to completed:

1. Read .claude/autoimprove/log.md
2. Find the last **Status:** IN_PROGRESS line
3. Replace with **Status:** COMPLETED (N/N)
4. Write the file back

# Remove any remaining experiment worktrees
for wt in .claude/autoimprove/worktrees/experiment-*; do
  git worktree remove "$wt" --force 2>/dev/null
done
git branch | grep "autoimprove/experiment" | xargs git branch -D 2>/dev/null
rm -rf .claude/autoimprove/worktrees

Print a final summary table:

━━━ Session Complete ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📊 Score: BASELINE → FINAL (+/- DELTA)
🔁 Iterations: N total — X kept ✅, Y discarded ❌
📝 Merged commits:
   • abc1234 autoimprove(001): description
   • def5678 autoimprove(003): description
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Universal Improvement Areas

Rotate through these (add language-specific ones from .claude/autoimprove/config.md):

• Type safety — fix type errors, replace any/interface{}/untyped constructs
• Error handling — unhandled promises, bare catch {}, swallowed errors
• Dead code — unused imports, variables, unreachable branches
• Code duplication — extract repeated logic (3+ occurrences) into shared utilities
• Naming & readability — cryptic names, functions over ~50 lines
• Performance — N+1 query patterns, missing memoization, unnecessary allocations
• Security — hardcoded secrets, missing input validation, unguarded auth routes
• Tests — add a test for the most critical untested function, fix flaky tests

Safety Rules

• Main branch is never modified until a winning experiment is explicitly squash-merged
• Never modify lock files, generated files, migrations, .env — in any worktree
• Never run deploy, publish, or push commands
• If the same area fails 3 iterations in a row, skip it and note in the log
• After 10 iterations, pause, clean up worktrees, and wait for human review
• On any unexpected error: run the worktree skill's cleanup step, then stop and report