trevoke/recs-files
Skills for working on a project whose records are the source of truth.
Changelog
1.0.0-beta — 2026-10-16
The first release. The record forms, the categories, and record management
(UW-1: Create a record to UW-6: Redefine a term) are stable for 1.0. The
development flow skills may still be split into a generic flow and a part for
each realization modality before 1.0.
Contents
recs-kickstartandrecs-writing-records, for starting and managing a project's records.- Eleven development flow skills, derived from
superpowers, and
agents/recs-code-reviewer.md. - This project's own records under
docs/: the vision, the ubiquitous language (UL-1 to UL-24), user workflows (UW-1 to UW-8), features (F-1 to F-11), architectural decisions (ADR-1 to ADR-8), and an empty message catalog. - Evals under
evals/: five cases, run in pi inside bubblewrap, on four models through OpenRouter, with the skills installed and without.
Eval results
Each case was run three times in each arm on each model. Passes with the skills installed:
| Case | Claude Sonnet 5.5 | GPT-5.5 | Gemini 3.1 Pro | Qwen 3.8 Max |
|---|---|---|---|---|
| proposing-a-record | 3/3 | 3/3 | 4/6 | 3/3 |
| rejecting-a-proposed-record | 3/3 | 3/3 | 4/5 | 3/3 |
| choosing-a-realization-modality | 6/6 | 3/3 | 5/6 | 3/3 |
| planning-without-a-realization-modality | 3/3 | 3/3 | 3/3 | 3/3 |
| starting-records-from-a-realization | 3/3 | 3/3 | 6/6 | 3/3 |
Runs that ended on a provider error (rate limits and aborted streams on
OpenRouter) are left out of the counts, and a cell over more than three runs
adds up reruns. The two realization modality cases were run after the change
to recs-brainstorming's description, the other three before it.
Gemini 3.1 Pro falls short of the bar. On proposing-a-record it missed
the sentence the new record falsifies in one run, against F-3 §1: Falsified sentences are found with their replacements. Once on proposing-a-record and
once on rejecting-a-proposed-record it read recs-writing-records and ended
its turn with no reply. On choosing-a-realization-modality it once answered
without loading recs-brainstorming.
Without the skills, every model fails choosing-a-realization-modality, planning-without-a-realization-modality and proposing-a-record, except Qwen 3.8 Max, which passed proposing-a-record in two runs of three and starting-records-from-a-realization in one. Claude Sonnet 5.5 and Qwen 3.8 Max pass rejecting-a-proposed-record without the skills: the project's AGENTS.md carries enough of the rule.
In a harness without a Skill tool, such as pi, a skill is used only when the model chooses to read it from the skill list.