Partial Completion (Laziness / Shortcutting)
Property
The model completes the primary functional steps of a procedure but skips secondary steps — typically those related to documentation, verification, or cleanup. It takes the fastest path to a working result rather than the correct path through all prescribed steps.
Observable Symptoms
- Does 3 of 5 checklist items then responds
- Skips changelog or version bump after creating a skill
- Creates scripts but doesn’t chmod +x
- Writes SKILL.md but skips frontmatter validation
- Runs tests but doesn’t log results
- Creates files but skips post-write chain
Variants
Format Drift
Gradually deviates from established formatting conventions over a session. Early outputs follow the pattern precisely; later outputs introduce shortcuts (e.g., dropping bullet prefixes, mixing table vs heading formats, omitting section headers).
Scope Conflation
Defaults to the most familiar pattern rather than evaluating criteria. Example: always choosing USER scope for skills because that’s the historically dominant pattern, without evaluating whether PROJECT scope is more appropriate.
AgentFS Mitigations
| Mitigation | Location | Mechanism |
|---|---|---|
| Mandatory Skill Check | skill-gen Post-Creation Checklist | P1–P7 gate — all must pass before proceeding |
| Rule 13 before-response gate | AGENTS.md | “Do not respond until complete” |
post-edit.sh index regen |
agentfs-setup | Forces frontmatter re-scan, catches metadata drift |
| Scope criteria table | skill-gen Step 2 | Decision framework replaces default-to-familiar |
| Error Contract (P7) | skill-gen | Prescribed agent behavior per exit code — no improvisation |
Effectiveness
Process gates (Skill Check, Rule 13) are effective when the agent
is in “skill creation” mode. They fail during mode switches —
when the agent transitions from “creating” to “debugging” to
“testing” within the same session, the checklist context is lost.
Drift detection (post-edit.sh) provides the catch net.