Second Look

New model. Old work. Show what actually got better.

Preserve the original problem. Spend a small budget revisiting the work that matters. Compare a patch, an intent-first revision and a fresh implementation against the same acceptance checks—with screenshots, regressions, tokens and cost.

Live demo · Quickstart · 한국어 · Releases

MIT · Python 3.11+ · macOS / Linux · 0.6 beta

User pilot: help decide whether this beta earns a second use. Try one real task and report adoption, rejection or setup friction. Participate · Decision rules, review on October 23. Feature expansion is on hold while we test actual use; the existing release remains available.

Try it with zero model calls

python3 -m venv .venv
source .venv/bin/activate
python -m pip install 'https://github.com/mandu5/secondlook/releases/download/v0.6.0/secondlook_revisit-0.6.0-py3-none-any.whl'
python -m playwright install chromium
secondlook doctor
secondlook demo --offline --output ./first-look

Open first-look/report.html in your browser. The deliberately flawed feedback app passes 3/8 checks; its hand-authored reference repair passes 8/8. This exercises the real browser evaluator without a model account. It is a synthetic demonstration, not a model benchmark. On Linux, browser system libraries may need python -m playwright install --with-deps chromium.

Give your own work a second look

Simple text requirements can be captured without writing JSON:

secondlook init ./my-app --files index.html \
  --intent 'Welcome readers and retain the documentation link' \
  --expect-text 'role=heading[level=1]' 'Hello' \
  --preserve-text 'role=link[name="Documentation"]' 'Documentation' \
  --output ./capsule.json
secondlook probe ./capsule.json --output ./baseline

Supply your requirements and selectors. Intent starts as a draft; review it and the checks before confirming. Add a separate requirements file for a clean-slate comparison. Complete own-project walkthrough.

Workflow	What the candidate sees
B · Patch	Original purpose and selected existing source
C · Reframe + patch	Purpose first; existing source only after an independent brief
D · Rebuild	Independent requirements and read-only data, without old writable source

Actual execution supports trusted static HTML/CSS/JS. Generate automatically with the tool-free Claude Code CLI adapter, or import responses from the AI tool you already use. Backends, arbitrary repositories and Python execution are not supported. Other problem types can be preserved in the library and stay parked.

Use your existing AI tool — no new API account

secondlook prepare ./capsule.json --mode ordinary --output ./request
# Give only request/request.md to your AI tool in a fresh conversation.
# Save its JSON response, then compare locally:
secondlook assess ./request --response ./response-a.json --label 'My chosen model' \
  --output ./comparison

Repeat --response and --label to compare up to four candidates. Use --mode rebuild with independently confirmed requirements to omit old writable source. The same frozen browser checks evaluate every candidate. These commands make zero model calls; external generation may have a cost. Model labels and optional usage records are owner-reported, and missing cost stays Unknown. Complete walkthrough and limits.

Revisit selectively

The problem library preserves original sources and intent across projects. Catalog observations can prepare queues for explicitly mapped candidates. Selection explains capability matches, declared impact/cost, missing adapters and previous attempts.

• Free baseline checks skip already-satisfied tasks.
• Identical task/source/checks/model/workflow/harness/trial attempts are skipped.
• Source projection can reduce bytes of context for large single-file apps.
• Bounded runs account for failed calls too. Unknown cost stops further calls.
• Original files remain intact; adoption is manual.

These are specific waste-avoidance mechanisms, not a measured universal token savings percentage. A catalog change does not automatically call a model.

Share the evidence

secondlook share ./first-look/result.json --output ./shareable
# Screenshots are opt-in because they contain visible page content:
secondlook share ./first-look/result.json --output ./visual-share \
  --include-screenshots --include-intent --title 'Feedback search and export'

Open or host index.html; keep summary.json beside it. Default exports contain metrics, generic check labels and recognized versioned model IDs. Source, prompts, raw receipts, private paths and detailed errors are excluded. Explicitly included text/images still need your review. Nothing is uploaded by this command.

What we have measured

Two existing project artifacts reached 6/9 → 9/9 and 3/7 → 7/7 checks with all three workflows. Rebuilding cost least on one; patching on the other. An already-passing third artifact made zero model calls. One initial checkset wrongly penalized a valid new DOM structure; that trial remains disclosed and separate from the corrected, newly frozen protocol.

The two interpretable trials used 8 calls, 46,812 tokens, $0.2544832 API-equivalent. Including the confounded trial: 12 calls, 77,088 tokens, $0.4133206. All used claude-sonnet-5-5; the old artifacts' generating models are unknown. These are artifact/workflow comparisons, not causal model-upgrade evidence or proof of general superiority. Protocol and all costs.

Promptfoo, LangSmith, Spec Kit and other projects cover important parts of this space. Second Look explores the connection between preserved intent, revisit selection and implementation alternatives. Prior art · Updated research and v0.6 decisions.

Contribute

git clone https://github.com/mandu5/secondlook.git
cd secondlook
python3 -m venv .venv && source .venv/bin/activate
python -m pip install -e '.[dev]'
python -m playwright install chromium
python -m pytest -q

Contributing · Security & data boundaries · CLI reference · Clean-slate protocol · Changelog · Roadmap

Live-run costs are CLI-reported API-equivalent estimates, not subscription invoices. Imported costs and model labels are owner-reported, not provider-verified. Provider budget stops can overshoot by an in-flight response. Local browser network isolation is not an OS sandbox for hostile code. Keep private reports private. No automatic adoption, telemetry or background model runs.