claudetools/session-logs/2026-06/2026-06-26-howard-skill-first-routing.md

User

• User: Howard Enos (howard)
• Machine: Howard-Home
• Role: tech

Session Summary

Howard raised a recurring failure: when asked to "put in billing for a company" in Syncro, the session was hand-rolling raw Syncro REST API calls (add_line_item via curl from memory) instead of using the /syncro skill, producing malformed tickets that Winter then had to fix. He wanted a hard, durable rule that when a request maps to a known skill, the skill is used — not improvised API calls.

Diagnosed the root cause as a routing-default failure, not a knowledge gap: the "act directly" bias plus knowing the Syncro API made hand-rolled curl the fast path, and nothing in the always-loaded CORE rules made skill use mandatory. The eight existing Syncro feedback memories describe correct API payloads but were being misread as license to free-hand the calls — notably feedback_syncro_billing.md §1 ("Bill with add_line_item directly") which meant "not the timer workflow," not "bypass the skill." Fixed by adding a Skill-first rule to CORE CLAUDE.md, creating feedback_skill_first_routing.md (+ index line), clarifying the ambiguous billing memory, and logging the correction to errorlog.md.

Howard then expanded the requirement: he doesn't want to name skills per-request at all. He wants a session to see a request, recognize which skill covers it (doing-skill: billing, coding, etc.), AND automatically run the matching verification/quality skills (code-review, security-review, the "make it sound human" pass, rule/standards checks, lint) before declaring work done — inferred from the request and expected end result, calibrated to stakes, and cheap (using Ollama Tier-0 for classify/prose passes). This was framed as "that's why we make the skills — to double-check things are how they should be."

Presented a three-part fix (A: CORE Definition-of-Done rule; B: on-demand .claude/SKILL_ROUTING.md map; C: deterministic Stop-hook backstop) and asked Howard for the enforcement tier and which work-types should auto-verify. He chose A+B now, hold C, with all four work-types auto-verifying (code, outbound/client-facing, Syncro/vendor, docs/wiki/memory). Implemented A+B: extended the CORE Skill-first bullet into a Definition-of-Done with a pointer to the map, authored SKILL_ROUTING.md (full request→doing-skill→check-skill table + stakes calibration + Ollama Tier-0 guidance), extended the memory and index line, then committed and pushed.

Key Decisions

• Fix the routing default in CORE, not just memory. The eight existing Syncro memories failed to prevent the behavior because only their index lines load each session; CORE rules bind every turn. The Skill-first + Definition-of-Done rules went into .claude/CLAUDE.md so they apply fleet-wide every session.
• Two-half rule: route + gate. (1) Map the request to a doing-skill; (2) gate the result with the matching check-skill before "done." Both inferred from the request, never user-named.
• Enforcement tier A+B, defer C (the Stop-hook backstop). A+B (CORE rule + on-demand map) are cheap and low-risk. The deterministic hook is the only thing that removes reliance on Claude's memory, but carries nag/friction risk ("don't cause more issues"), so it's deferred until the rule+map are proven. Turning it on later is a one-line change; the trigger to do so is catching a skipped gate.
• On-demand map file, not CORE bloat. SKILL_ROUTING.md holds the full table and is read only when a request maps to a skill or work nears done — keeps CORE lean and costs no tokens until triggered.
• Calibrate to stakes + Ollama Tier-0. Trivial edits skip heavy review; security/auth/outbound always get the full gate. Cheap classify/prose passes route to Ollama to keep token cost low (Howard's explicit constraint).
• Clarified the ambiguous billing memory rather than deleting it: feedback_syncro_billing.md §1 "directly" means "not the timer workflow," not "bypass the skill" — closed the misread that caused the original problem.

Problems Encountered

• Apparent "lost" git add — after staging the first batch of files, git status --short showed only the pre-existing ampipit submodule change. Cause: the machine's auto-sync had already committed the edits in commit 5bace24 ("auto-sync from HOWARD-HOME at 2026-06-26 11:40:19"). Verified via git log -- <file> and git show --stat HEAD. No data lost; confirmed ## main...origin/main in sync.
• Second batch timing — to avoid relying on the auto-sync timer for the A+B changes, committed and pushed them explicitly. LF→CRLF warnings on the two new/edited files are benign Windows line-ending normalization.

Configuration Changes

Modified:

• .claude/CLAUDE.md — added Skill-first key rule (use the skill, never hand-roll the API; Syncro billing always via /syncro); extended it into a Definition of Done (route to doing-skill, gate with matching check-skill before done; calibrate to stakes; Ollama Tier-0; pointer to SKILL_ROUTING.md).
• .claude/memory/feedback_syncro_billing.md — added clarifying paragraph: run billing THROUGH /syncro; §1 "directly" = "not the timer workflow," not "bypass the skill."
• .claude/memory/MEMORY.md — added/updated Syncro-section index line for feedback_skill_first_routing.
• errorlog.md — logged the correction (--correction).

Created:

• .claude/memory/feedback_skill_first_routing.md — the Skill-first + Definition-of-Done feedback memory.
• .claude/SKILL_ROUTING.md — on-demand request→doing-skill→check-skill routing map.

Credentials & Secrets

None discovered, created, or rotated this session.

Infrastructure & Servers

None touched. (Referenced only in routing examples: Syncro at computerguru.syncromsp.com; coord API http://172.16.3.30:8001/api/coord — no changes.)

Commands & Outputs

• bash .claude/scripts/log-skill-error.sh "syncro/billing" "hand-rolled add_line_item API calls from memory instead of using the /syncro skill; ..." --correction --context "rule=skill-first memory=feedback_skill_first_routing tickets=#32193,#32194" → [OK] logged skill error to errorlog.md (syncro/billing)
• git show --stat HEAD confirmed auto-sync commit 5bace24 captured the first batch (CLAUDE.md +11, MEMORY.md +1, feedback_skill_first_routing.md +18, feedback_syncro_billing.md +2, errorlog.md +2).
• Second commit (A+B): harness: add Definition-of-Done skill routing ... pushed to origin/main; git status -sb → ## main...origin/main.

Pending / Incomplete Tasks

• Option C (deterministic Stop-hook backstop) deferred — non-blocking hook that reminds when code was edited but no review ran. Turn on if/when a skipped gate is caught in practice. One-line settings.json change via update-config.
• Maintain .claude/SKILL_ROUTING.md as skills are added/changed (it is the durable map).
• Pre-existing projects/msp-tools/ampipit submodule change remains uncommitted — not part of this work; left untouched intentionally.

Reference Information

• CORE rules: .claude/CLAUDE.md (Skill-first + Definition of Done under "Key rules (always)").
• Routing map: .claude/SKILL_ROUTING.md.
• Memory: .claude/memory/feedback_skill_first_routing.md; related feedback_syncro_billing, feedback_syncro_preview_mandatory, feedback_psa_default_syncro, feedback_syncro_priority_type_format, feedback_impeccable_on_outbound, feedback_calibrate_effort_to_stakes.
• Syncro malformed-ticket precedents Winter flagged: tickets #32193, #32194 (and the priority/type-format memory).
• Auto-sync commit: 5bace24.