Trim /setup skill from 475 to under 400 lines by removing redundant content

[sebastientaggart]

Problem to Fix

The /setup skill is 475 lines — the longest skill in the project — despite being a one-time configuration wizard. Much of the length comes from duplicated label baselines, verbose config walkthrough entries that repeat an identical pattern longhand for each value, a skills reference table that duplicates the README, and profile selection with redundant follow-up questions.

Why it Matters

Longer skills increase the chance of instruction-following errors and make maintenance harder. A one-time wizard doesn't need to carry the same weight as frequently-invoked workflow skills like /start or /submit-for-review.

General Approach

Trim setup.md from ~475 to ~375 lines by:

• Removing the skills reference table (already in README)
• Merging the greenfield label guidance into Phase 2's existing label population flow
• Defining the Phase 3 config walkthrough pattern once, then listing values as a compact table
• Factoring out shared profile selection questions

No behavioral changes — the setup wizard should do exactly the same things, just with less redundant instruction text.

Complexity

Verification / QA effort: moderate
Requires running /setup against projects in each state (State 1, State 2, State 3) to verify behavior is preserved after trimming.

Acceptance Criteria

• setup.md is under 400 lines
• All three states (just checking it out, partial setup, fully configured) produce the same behavior as before
• No setup wizard steps are removed or reordered
• Phase 2 label population and greenfield guidance are consolidated into a single flow
• Phase 3 config walkthrough uses a defined pattern rather than repeating longhand for each value

[sebastientaggart]

Agent Implementation Notes

File to change: skills/setup.md (single file change, then sync.sh to regenerate)

Approach — four targeted cuts:

1. State 1 skills table (lines 57–65) — Delete the 10-line table. The sentence on line 66 already points to README.md. Compress Path A's three-layer explanation slightly.

2. Greenfield baseline guidance (lines 358–376) — Merge into Phase 2 (lines 279–339). Phase 2 already has the exact same label list and creation flow. Add the milestone guidance sentence from greenfield into Phase 2's flow after label selection. Delete the standalone greenfield section entirely.

3. Phase 3 config walkthrough (lines 343–447) — Replace the 5 longhand entries with:

   • A single pattern definition: "For each unset value: explain in one sentence, show example value, ask if they want to set it. If yes, show exact YAML change and confirm before writing. If skip, move on. Do not ask again."
   • A compact table: | Key | Profile | Description | Example | with one row per value
   • Keep PLATFORM_COMPLIANCE_NOTES and CONVENTIONS_NOTES iterate-until-approved behavior as a single note after the table, since those two require drafting rules from user input (unlike the others which are simple values).

4. Profile selection (lines 151–197) — Factor out the shared "What's your production branch?" and "What's your integration branch?" questions into a preamble. Each profile block then only lists which questions apply and which values to write.

Edge cases:

• The greenfield guidance has a condition (TICKET_LABELS unset OR <5 labels). After merging into Phase 2, this condition should gate the baseline recommendation within Phase 2's flow.
• PLATFORM_COMPLIANCE_NOTES and CONVENTIONS_NOTES have a draft-iterate-confirm cycle that differs from the simple ask-and-set pattern. Must preserve this distinction.

After editing: Run CodeCannon/sync.sh to regenerate all adapter outputs, then wc -l skills/setup.md to verify line count.

[sebastientaggart]

Resolution

The setup wizard instructions were trimmed from 475 to 371 lines by consolidating duplicated label guidance, replacing repetitive per-value walkthrough entries with a compact table, and removing a skills reference table that duplicated the README. The roadmap file was also removed since all items are now tracked as backlog tickets. No steps were removed — the wizard behaves identically with less redundant instruction text.

See #112 for full technical details.