Bug] Several references demand full-cost work the request doesn't need, or contradict SKILL.md

Open
#985 2 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

  • #1118 by @srobroek — closed without merging

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
58/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Active
Tech stack
javascript
Domain
tooling

Research direction

Compare the named files: skill/reference/new-work.md, critique.md, layout.md, typeset.md, harden.md, onboard.md, and overdrive.md against skill/SKILL.src.md, skill/reference/init.md, and crates/context/src/seed_text.rs. Start by checking each reported conflict and the relevant reproduction steps. Done means the references follow the expected behavior, with the changes limited to the source files as described.

Written by the indexing model from the issue text.

Description

bug P2 policy: approved

What happened?

Some references require their most expensive path on every call, including when the request is small or already decided, or contradict another rule in the skill:

  • new-work.md: step 4 makes concept-seed mandatory on every new or replacement world, "whatever the harness, the model, or the time pressure", and step 1 lets a brief's own picture take at most one of the seven candidates. SKILL.md says "the brief wins" and the roll output says "a user- or brief-pinned decision beats the roll, always". A brief that pins a world, palette or type direction still goes through the full derivation, the roll and the decision page, and its own picture can lose to a dealt challenger.
  • critique.md: project-specific personas read a ## Design Context section in CLAUDE.md/AGENTS.md. init no longer writes that section (it writes PRODUCT.md), so those personas never generate. Every critique also has to spawn two parallel sub-agents, and an inline run is "NOT permitted" even for a single small component.
  • layout.md, typeset.md: every call runs two isolated assessments, answers every question with evidence, scans, then answers every verify item and rescans, even for one component's spacing or one heading weight.
  • harden.md: it opens by scoping itself to the interface, then requires server-side validation and rate limiting "always", lists ignoring internationalization and forgetting offline as NEVERs, and asks for language, offline and old-browser test sweeps on every run.
  • onboard.md: "Every empty state needs" five parts, illustration and tutorial link included, directly above the type list that says a user-cleared state wants a light touch. Its library list names Popper.js (now Floating UI) and Intro.js and Shepherd.js with no note that both are AGPL-3.0 or commercial.
  • overdrive.md: "Expect multiple rounds" of browser iteration, against SKILL.md's bounded verification (one batched round, one fix batch, at most one more round). It also forces a response banner and uses "EXTRA IMPORTANT" / "You MUST" emphasis.

Steps to reproduce

  1. Compare skill/reference/new-work.md steps 1 and 4 with skill/SKILL.src.md ("The brief wins") and the A user- or brief-pinned decision beats the roll, always. line in crates/context/src/seed_text.rs.
  2. Compare skill/reference/critique.md (Project-Specific Personas) with the PRODUCT.md template in skill/reference/init.md, which has ## Users and ## Operating Context but writes no ## Design Context.
  3. Compare skill/reference/overdrive.md ("Iterate with Browser Automation") with the bounded-verification line in skill/SKILL.src.md.

Expected behavior

A pinned brief is honored without a tournament, and an open brief still gets one. Critique personas use PRODUCT.md. Isolated assessments, evidence passes and rescans scale with the size of the change. harden handles what the interface shows and reports server-side or locale gaps instead of mandating them. Each onboarding empty-state type gets what it needs, and the library picks are current. overdrive follows the skill's bounded verification.

Provider & environment

  • Provider: found by reading the skill source on main (778c8a7b7); not provider-specific.
  • Provider version: n/a
  • OS: macOS 26 (arm64)

Additional context

Each item is a separate commit on my branch, so you can take some and leave others. The references stay about the same size (about 1% more bytes in total). The saving is in the work they no longer require on every call: a pinned new-work run skips the roll, whose recorded concept-seed output alone is about 3k tokens, plus the decision page and its comps, and a small-target critique skips two sub-agent spawns.

AI assistance: prepared with an AI coding agent (oh-my-pi harness, Anthropic Claude) under my direction; I reviewed it.

Willing to work on a fix?

Yes. I have a branch ready (source files only, no generated output, no golden changes) and will wait for maintainer approval before opening a PR.

Dominant language
Rust
Stars
111k
Forks
6.7k
Avg merge
20h 3m
Merged PRs (30d)
344

Getting set up

  • No Dockerfile or Docker Compose file
  • Has a pull request template
  • No contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from pbakaus/impeccable

All issues in pbakaus/impeccable

Similar issues

More Rust issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.