No repository context
Run each task from its historical base commit with repository instruction files hidden.
See what you get · no signup
A real brief, before you try your repo.
This snapshot was prepared from the public astral-sh/uv repository on September 28, 2026. It shows the actual output of the planner. No agents were run, and no performance improvement is claimed. The project is not affiliated with Sharper Than AI.
Prepare your own free brief →Example snapshot · not executed
astral-sh/uv
6Historical tasks
1Instruction files
608Estimated tokens
34Context lines
Pull requests are historical outcome anchors, not finished prompts. Review each task description and define its evaluator before running the experiment.
01
Controlled comparison
Change one variable
Current context
Repeat from the same base commit with the detected instruction surface restored.
Held constantAgent and model versionTask promptRepository base commitTool permissionsStep and token budgetEvaluator and test commands
02
Candidate task set
Anchor tests in real merged work
PR #22051
Revert "Batch HTTP cache writes in blocking tasks"
Start at 4d3ce16 · compare against 07f4df6
PR #22050
PR #21951
Normalize requirement declarations in lockfiles
Start at 7a3bc07 · compare against 13d1219
enhancementpreview
PR #22047
Downgrade some of reqsign's transitive deps
Start at 7a3bc07 · compare against a5990bd
internaldependencies
PR #22033
Avoid panics for unknown managed Python implementations
Start at 8373633 · compare against 8ec0039
bug
PR #114
Store all distributions rather than compatible wheels
Start at 5b046a8 · compare against 2799fe6
03
Detected treatment
Instruction surface
AGENTS.mdAGENTS.md
Repository-wide
608 tokens
04
Static review
Instruction health checks
No obvious static issues found
The automated checks found no oversized files, thin instruction files, duplicated long rules, or broken relative Markdown references. Runtime testing is still required.
05
Outcome scorecard
What the runner must measure
Verified completion
Repository tests and task-specific assertions pass.
Regression count
Previously passing behavior remains intact.
Cost
Input/output tokens and provider spend per successful task.
Time
Wall-clock duration and number of agent steps.
Exploration
Files read or modified outside the expected change area.
Instruction adherence
Concrete applicable rules followed or violated.
Put this brief to work
Replay it with the free local runner, review the evidence, and keep noisy results inconclusive.
This is an experiment brief, not a benchmark result.
Sharper Than AI has prepared reproducible inputs from public repository evidence. It has not run a coding agent or claimed that the current instructions help.