Your work. Your next useful decision.
Make your next attempt count. Sharpen your prompt, shape how AI works with you, and check what changed in your own work.
On your laptop · Claude Code or Codex · No account needed
Pick a problem. Make your goal, prompt and acceptance check clear.
Review a scoped working agreement, then make a deliberate next attempt.
Check your decision and outcome. Better, worse, or unchanged: keep the lesson.
Curated public work
Real work, published in the open. Find a useful technique in a notebook, guide, report or project, then try it in your own work.
Browse 28 curated public examplesPublisher names credit the original sources. They do not imply Better Loop membership or a shared assessment.
Code & software
Ronacher published a tolerant XML parser after earlier parsing approaches failed. He describes directing the initial implementation, reviewing it, requesting further fixes and an extensive test suite, and releasing the resulting library.
Analysis & finance
The notebook applies a 60-month rolling CAPM to technology-industry excess returns using Ken French’s factor and industry data. It displays coefficient tables, confidence-interval plots, and an expanding-window example.
Research & strategy
The guide projects Chicago community areas and grocery locations into a metric coordinate system, buffers stores by one kilometre, and displays intersection and difference maps. It defines coverage geometrically.
Math & science
The notebook joins Minnesota household and county data, fits pooled, unpooled and hierarchical radon models, and displays posterior diagnostics and comparisons. The partial-pooling plots show how estimates change with county sample size.
Writing & design
OpenAI works through furniture images from an Amazon product dataset, generating descriptions and then concise captions. Printed examples show how the product title identifies the intended item when an image contains several objects.
Operations & education
Software Carpentry’s lesson reads Gapminder country data and displays labeled GDP time-series and scatter plots. It shows table transposition, legends and file export, alongside learner exercises and accessibility guidance.
Find a place to begin
Reconcile two totals. Challenge an answer. Give a draft a clearer brief. Start with a familiar situation and one check that matters.
Explore the practice boardsAuthored practice briefs. Choose routine, moderate or complex work; no participant counts or earned scores are implied.
Better Loop project runs
Start with our public evaluations: a repair that worked, guidance that didn’t improve the checked result, and host behavior with limits. These are project case studies, separate from community contributions.
Would extra guidance help identify when an approval remains valid?
Both attempts passed 12/12 cases and 3/3 scope checks. Guidance used 364 more reported tokens (about 1.05%).
A more elaborate prompt needs to earn its place. Keep the same acceptance check when you try it.
Quality: No measured quality gain on the frozen deterministic checks.
Resources: 34,617 → 34,981 reported tokens. All reported main and auxiliary model entries, including cache use, counted once. Shared overhead, cash billing and human effort unknown.
Comparison: One task, one pair, baseline first. Same frozen public code, structured-answer task and bl-approval-binding-judge-0.1 evaluator. Reported model entries include the auxiliary classifier.
Difficulty: Rubric-estimated in the frozen registration.
Exact reported model entries: claude-opus-5[1m], claude-haiku-4-5-20251001. These historical labels do not assert current availability.
Not counterbalanced or independently held out. No full-skill, human-improvement, speedup or savings claim.
Inspect the recorded measurementsCould a validator reject unsupported human-action claims and still accept missing evidence?
Frozen checks passed 16/24 before and 24/24 after. Eight known failures were fixed; no case regressed.
Write a test for the mistake and a control for what must keep working. Missing evidence can be the correct answer.
Quality: All 24 decisions, six positive controls and ten negative controls met the prespecified floor.
Resources: No models invoked. Coding effort, model resources and cash cost were not measured. Execution duration is not a before/after speed comparison.
Comparison: Retrospective comparison of evidence draft.1 and draft.2 on the same known authored probes. Evaluator bl-evidence-attribution-repair-0.2; before then after.
Difficulty: No difficulty rating recorded for this repair evaluation.
Known, non-blinded cases. The initial infrastructure attempt evaluated no cases; the corrected retry is not replication. This measures a software repair, not a person's ability or model efficacy.
Inspect the recorded measurementsThese different tasks, versions and evaluators cannot rank models, hosts or people. Public project measurements do not enter community cohorts, count as challenge completions or establish improved human ability.
Explore Codex, Claude and software project runsLearning from real work
Read what changed, what was checked, and what happened. A useful negative result can save you an unhelpful next step.
Keep an unsuccessful attempt visible. Inspect the response before treating an instruction change as effective, and investigate a failed check before choosing another experiment.
This task comparison · baseline 100
Reported · rounded to five points
higher is better
The task and source text stayed the same. Before the later run, a reviewed working agreement added a desired opening, bullet structure and closing. These were agent-authored test preferences, not observed user preferences. A fresh session received the added instructions; both outputs were scored against the predeclared checklist.
Evidence limits: One before-and-after pair on a public writing task. Source-fact checks were reviewed manually, without blinding. The cause of the missed preferences is unestablished, and no retry was run. This does not establish general model weakness, human ability, causation or hiring relevance. Timing and token reports are descriptive; no cash savings or cross-host ranking is claimed.
creating content · Self-reported evidence · limited coverage
See the change and the evidenceMake project preferences testable, start a new session after applying them, and inspect the next response. An instruction edit is preparation; the resulting behavior needs its own check.
This task comparison · baseline 100
Reported · rounded to five points
higher is better
The task and source text stayed the same. Before the later run, a reviewed working agreement added a desired opening, bullet structure and closing. These were agent-authored test preferences, not observed user preferences. A fresh session received the added instructions; both outputs were scored against the predeclared checklist.
Evidence limits: One before-and-after pair on a public writing task. Source-fact checks were reviewed manually, without blinding. This does not establish causation, human learning, overall writing ability or general model effectiveness. Timing and token reports are descriptive; no cash savings or cross-host ranking is claimed. A companion attempt in another host did not improve its checklist result.
creating content · Self-reported evidence · limited coverage
See the change and the evidenceWhen changing a validation rule, test examples that must be accepted as well as examples that must be rejected. Keep previously passing controls and retain unsuccessful attempts.
This task comparison · baseline 100
Reported · rounded to five points
higher is better
The comparison ran the earlier and repaired public code against the same frozen acceptance cases. Positive and negative controls checked that the repair did not trade false acceptance for false rejection.
Evidence limits: This is a retrospective comparison of actual public code on known authored probes, not an unseen evaluation. The normalized index describes passed checks on that fixed set. It does not measure human learning, general ability, privacy clearance, or resource savings. Model and human effort were not measured.
diagnosing error · Self-reported evidence · limited coverage
See the change and the evidenceYour practice comes first
Keep assessments, before/after work and your return checkpoint on your laptop. Private coaching stays useful even if you never publish.
Raw work stays out of Better Loop. Your configured model provider may process the evidence you select.
Build your practice →A later human decision, a clear check and an honest comparison of the work you selected. Copying prompts or applying preferences is preparation, not measured improvement. Results may be negative or inconclusive. Public evidence is self-reported; it does not establish general ability or guaranteed gains. See the measurement approach.