Opus 4.7 (Claude Code) ignores explicit numbered steps in a user-authored skill, substitutes its own abbreviated process

Model: claude-opus-4-7 (1M context), Claude Code CLI
Platform: darwin (macOS)

Summary

A user-authored skill (slash command) defined an explicit, numbered step list — one task per step — and the model did not follow it. Instead of executing the steps verbatim, the model substituted its own abbreviated process: it created the wrong set of tasks, dropped mandated steps, conflated others, created them in the wrong place, went out of order, and continued past an explicit approval gate without being told to. This defeats the purpose of an authored skill: a clearly documented, step-by-step workflow is being treated as loose guidance.

The skill

The skill's "Progress Tracking" section enumerated nine step-tasks to create up front:
Step 0, Step 1, Step 2, Step 3, Step 4, Step 5, Step 6, Step 6.5, Step 7 — each with a defined purpose, and an explicit gate ("present coordination graph; do not spawn until approved").

What the model did instead

1. Created six tasks, not nine. Dropped Step 3 (create granular agent tasks), Step 5 (enforcement loop), and Step 6.5 (re-review fix-ups). Conflated Step 4 and Step 5 into one task.
2. Created the tasks before provisioning the team, so they landed in the wrong task list.
3. Prepared to proceed out of order (presented the Step 2 gate while Step 3's prerequisite tasks did not exist).
4. When told "redo from scratch," recreated the nine tasks but immediately marked early steps completed/in-progress instead of starting from a clean pending state.
5. Continued narrating/advancing through subsequent steps without the user saying continue, past the documented approval gate.

Expected

When a skill enumerates explicit, numbered steps, execute them verbatim and in order: create exactly the tasks the skill specifies, honor each gate, and stop where the skill says to stop. Do not substitute an abbreviated or reordered process.

Impact

• Authored skills/workflows become unreliable — the user cannot trust that a documented process will run as written.
• Forces repeated correction and re-issuance, wasting the user's time.
• Undermines the value proposition of skills entirely ("why have skills if they don't get followed").

Notes

Not a knowledge gap: the full skill text was in the model's context at invocation. The failure is the model substituting its own process for an explicit, in-context, numbered one — an instruction-adherence / determinism problem for structured multi-step skills.