About this skill

Problem: why “make it better” rarely converges

When an AI draft is close but not quite right, the usual fix is another pass. Without a goal, rubric, and rollback rule, edits become local churn: one sentence improves while another weakens, or a bug fix makes the code harder to read. autoresearch-lite targets quantifiable, iterative refinement, not open-ended brainstorming. It defines “better” first, then enters a traceable revision loop.

How it works: scoring-guided iteration

The user supplies three inputs: the target content, the evaluation criteria, and optionally a round limit. The skill first builds a baseline: it reads references/scoring-templates.md, locks the rubric, scores the original as iteration #0, and records an anchoring rationale for every dimension. These rationales point to concrete sentences, structure, or parameters, preventing scores from drifting upward without real improvement.
Each round follows a fixed loop: review the best version and history, choose one edit direction, make one focused change, then rescore. Directions include removing redundancy, replacing vague claims, restructuring, adding evidence, or adjusting tone. Improving scores are kept; regressing scores are rolled back. Stalls trigger broader changes or a stop signal. In workspace mode, state is persisted under autoresearch/ in baseline.md, best.md, state.md, and log.md, so interrupted sessions can resume. For code, if commands can run, it can also save the file, run tests, and use observed output as evidence for robustness and performance.

Boundaries and notes

It fits tasks with a clear standard: business copy, prompts, technical plans, code quality, and titles. It does not fit one-off fixes, unmeasurable goals, or tasks that mainly need fresh inspiration. Scoring remains partly subjective; the user can correct the anchoring rationale at any time. Factual changes are marked [needs confirmation] rather than invented to inflate scores. If the baseline is already strong, progress stalls, or edits involve real names, institutions, or data, it should pause and ask instead of publishing, committing, or sending anything.

Screenshots

Use Cases

Revise a technical proposal when the logic feels weak, using scored dimensions and keep-or-discard logs.
Improve a Python API by running tests first, then making single focused edits to readability and robustness.
Iterate on an unstable Prompt against a target score, keeping valid changes and delivering the final version.
Refine a long report over 2,000 words by locating weak sections and iterating them before a final pass.

Best For

Business copywriter who needs to turn “not engaging enough” into scored dimensions for iterative revision.
Engineer maintaining agent code who wants readability, concision, robustness, and test-backed changes.
AI app developer iterating prompts who needs valid edits kept and regressions rolled back.
Skill author continuing from a luban nine-dimension baseline who wants inherited anchors and an acceptance summary.