Repository navigation
Conversation
Apply the Claude 5-generation context-engineering rules: progressive disclosure over upfront detail, a single source per instruction, and expressive descriptions instead of worked examples. - Strip <examples> blocks from six agent descriptions and rewrite them to route on their own. Those blocks were concatenated into every request, costing ~2,400 tokens per turn for agents dispatched by name anyway. - Merge when_to_use into description so the trigger surface has one source. - Extract the procedures duplicated across skills into shared references: feature-branch.md (brainstorm, plan, debrief) and review-dispatch.md (build, hotfix, review). - Demote run-time detail out of SKILL.md bodies: commit-autonomy.md and ship-gate.md (build), plan-authoring.md (plan), issue-previews.md (debrief). - Drop trailing Important / Key Principles blocks that restated the body, moving each load-bearing line to the step it governs. - Document the KEEP / DEMOTE / DELETE rubric and the agent-description rule in CONTRIBUTING.md, with a pointer from AGENTS.md.
Second pass over the same surface, run through the claude-api skill's
prompt-audit workflow against the Claude 5 generation. The first commit
restructured where instructions live; this one fixes the instruction text
itself.
Under-described (the fix is more text, not less):
- Expand four agent descriptions that stated neither what the agent
returns nor when not to use it — pr-readiness, best-practices-research,
official-docs-research, and user-flow-analysis. These are the four the
first commit never touched, since only the other six carried <examples>.
- Give skills/hotfix a trigger surface. It was the one skill with no
trigger phrases at all, having never had a when_to_use field to merge.
Dated prompt text:
- Drop role inflation that substituted for context ("elite", "seasoned
Senior Engineer", "knows all the ins and outs"). Role lines that pair
the role with its reason are left alone.
- Remove the "Your mission" list in user-flow-analysis that the section
headings below it already restate, plus "Be exhaustively thorough" —
current models are proactive by default.
- Normalize caps emphasis: "MANDATORY Deprecation Check" and four
"DO NOT proceed until..." gates.
- Fix a drifted duplicate: hotfix capped PR titles at 70 characters,
create-pr and conventional-commits at 72.
Left alone deliberately: trigger-phrase enumeration in skill descriptions
(routing text, and changing it without a trigger eval is guessing), and
the YAGNI/testing prohibitions, which encode VGV policy from AGENTS.md.
Claude 5 models follow instructions literally and default to medium effort, so the prompts and frontmatter needed retuning. - Pin effort on every skill and agent; raise the code-review agents to high and lower /review to medium - Drop emphatic wording (ruthlessly, zero tolerance, MANDATORY) that made reviewers over-flag - Brief subagents fully: dispatch now passes the subject, base branch, merge-base, and changed files, since agents start without context - Review agents report every finding with a confidence level; callers check each finding against the code before fixing it - Replace the post-ship Done menu in /build and /hotfix with a closing summary - Document model and effort settings and instruction-writing guidance in CONTRIBUTING.md Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Author
|
Sorry for that 🤦 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
Retunes Wingspan's skills and agents for the Claude 5 model generation. These models follow instructions closely and literally, apply strong wording more broadly than intended, and in Claude Code start at
mediumeffort. The branch makes three passes over the same files.1. Structure (
refactor: prep skills and agents for claude 5gen models)<examples>blocks from six agent descriptions and rewrites those descriptions so they route without them. The blocks were sent with every request (about 2,400 tokens per turn), even though skills already dispatch these agents by name.when_to_useintodescription, so each skill's triggers live in one field.feature-branch.md,review-dispatch.md), and moves run-time detail out ofSKILL.mdbodies intoreferences/.CONTRIBUTING.md.2. Instruction text (
refactor: apply prompt-audit findings to agents and skills)/hotfixtrigger phrases.MANDATORY,DO NOT proceed)./hotfixcapped PR titles at 70 characters while/create-prallowed 72.3. Behavior and settings (
feat: tune skills and agents for claude 5 models)efforton every skill and agent. The code-review agents run athigh./reviewdrops tomedium, since its agents do the judging.high/medium/lowconfidence rather than deciding on their own what to leave out. The consolidated report gains a Confidence column, and/buildand/hotfixcheck each finding against the code before fixing it./buildand/hotfixnow end with a summary that stands on its own, replacing the "Done" menu that had only one option.CONTRIBUTING.md.What users will notice: review reports have a Confidence column and may list more findings, each graded;
/buildand/hotfixno longer end on a "Done" prompt.Validation
claude plugin validate .passes with two warnings, both from before this branch: the rootCLAUDE.mdisn't loaded as plugin context, andhooks.jsonleaves${CLAUDE_PLUGIN_ROOT}unquoted.hooks/test_recommend_plugins.sh: 42 passed, 0 failed.xhighis added to the cspell dictionary.Type of Change
feat)fix)refactor)docs)ci)chore)🤖 Generated with Claude Code