| name | drivemind |
| description | Use when meaningful agent work risks degrading or settling below the model's real capability: long tasks, pressure, goal drift, boundary drift, fake retries, weak blocker diagnosis, shallow outputs, continuity breaks, or work that should leave reusable residue. Do not use for simple one-shot tasks where direct execution is enough.
|
DriveMind
DriveMind is a bounded execution skill.
It improves hard agent work by protecting three things:
- Execution integrity โ keep the work aligned, bounded, and recoverable.
- Execution ceiling โ force one stronger concrete move when the current result is too shallow.
- Experience compounding โ leave residue that makes the next run stronger.
If DriveMind only adds ceremony, stop using it.
Activate When
- the task is long-running, high-value, blocked, or likely to cross session boundaries
- the user asks to keep pushing, stay steady, not stop too early, or force a better path
- the work risks goal drift, boundary drift, continuity decay, stuck degeneration, capability underuse, or closure failure
- the next action may involve risk, authority, deletion, publication, credentials, external messages, or irreversible impact
- the task should leave behind a review, lesson, next-time rule, handoff, memory, or SOP residue
Do Not Activate When
- the answer is simple and direct
- no continuity, boundary, stuckness, or compounding risk exists
- the skill would produce process without changing the next action
Main Path
1. Detect The Failure Mode
Choose the primary risk:
- goal drift
- boundary drift
- continuity decay
- stuck degeneration
- capability underuse
- closure failure
If none applies, continue normally without DriveMind.
2. Stabilize The Thread
State the minimum useful version of:
- current objective
- what matters most
- blocker, uncertainty, or boundary
- next action
Keep this short. Stabilization is not a status performance.
3. Protect The Boundary
If authority, risk, irreversible impact, or external representation is in play, stop and check the boundary before acting.
Read references/boundary-preservation.md.
4. Recover From Stuckness
If the work is blocked, diagnose before pushing harder.
Read references/stuck-recovery.md.
5. Raise The Execution Ceiling
If the work is moving but weak, force one stronger concrete move:
- inspect instead of guessing
- test instead of assuming
- compare instead of accepting the first path
- isolate the blocker instead of naming it vaguely
- produce the artifact instead of proposing it
- preserve the next executable action instead of writing a generic summary
Read references/execution-ceiling.md.
6. Preserve Continuity
If the work may pause or resume later, preserve objective, state, blocker, boundary, and next executable action.
Read references/continuity-preservation.md.
7. Close With Residue
If the work was meaningful, leave the smallest residue that will improve the next run.
Read references/closure-compounding.md and references/residue-selection.md.
Output Pattern
When DriveMind is active, prefer:
- objective or judgment
- what is degrading or too weak
- stronger next action
- boundary or escalation point if any
- residue when useful
Use natural prose. Do not expose the framework unless it helps the user decide or continue.
Reference Spine
| File | Use when |
|---|
references/drift-prevention.md | objective or thread coherence is weakening |
references/boundary-preservation.md | authority, risk, or irreversible action is in play |
references/stuck-recovery.md | blocked work risks fake motion |
references/execution-ceiling.md | the result is below the model's reachable capability |
references/continuity-preservation.md | work may pause, hand off, or resume later |
references/closure-compounding.md | the task should leave a useful ending |
references/residue-selection.md | choose the right residue form |
references/mode-guide.md | choose normal, execution, or intensive mode |