Menú
Diagnose and recover from dissertation agent failures such as false runs, repeated tool loops, stale assumptions, context drift, wrong-window behavior, or environment-state mismatches before retrying.
Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.
Basado en la clasificación ocupacional SOC
Design and review active learning implementation support for university teachers, including classroom activity structures, preparation workflow, student participation, assessment alignment, and AI-agent support points.
Turn dissertation evidence into an AI agent design specification for supporting active learning implementation, including user scenarios, functions, interaction flow, boundaries, failure modes, and design rationale.
Synthesize confirmed co-design, design-elicitation, prototype-feedback, or participant-generated design outputs into requirements, design principles, prototype changes, participant contribution maps, and design decision logs.
Design specifications for dissertation figures such as conceptual frameworks, concept-card/interview process diagrams, AI-agent workflow diagrams, adoption-condition models, and findings maps.
Maintain a lightweight dissertation research wiki with project memory, literature notes, methodology decisions, supervisor feedback, design rationale, and open questions.
Audit an AI agent prototype or design concept for usability, pedagogical fit, teacher feedback, active learning support, evidence quality, and next design iterations.
| name | dissertation-agent-self-debug |
| description | Diagnose and recover from dissertation agent failures such as false runs, repeated tool loops, stale assumptions, context drift, wrong-window behavior, or environment-state mismatches before retrying. |
Use this skill when the agent appears stuck, repeats actions, claims work without evidence, uses the wrong window role, misreads local state, or drifts from the user's actual dissertation maintenance task.
Adapt ECC agent-introspection debugging to this dissertation system. The goal is controlled recovery: capture the failure, classify the cause, verify the real state, take the smallest safe action, and record a prevention note if needed.
Record:
Use:
Failure capture:
- Task:
- Symptom:
- Last verified step:
- Suspect assumption:
- Files/tools:
- Risk level:
Classify the issue:
| Pattern | Likely Cause | Check |
|---|---|---|
| false completion | answer based on memory, not local evidence | inspect actual file/path/output |
| loop / no progress | repeated plan or repeated command | inspect recent actions and stop retries |
| context drift | old task or old window role dominates | restate current user request and window role |
| stale memory | TASK_STATE or wiki outdated | compare local files and latest task entry |
| wrong source | inferred facts used as confirmed facts | run source-first gate |
| tool state mismatch | browser, cwd, renderer, Obsidian, or config not as assumed | verify actual surface |
| surface mismatch | source layer changed but user-visible surface was not checked | compare file/git state with rendered page, app view, exported file, or public repo surface |
Take the smallest safe action:
research-wiki/TASK_STATE.md if the failure affects future workDo not claim that hidden state was reset unless an actual tool action proves it.
When the failure involves public release, Obsidian, browser, Word/PDF, Canvas, or GitHub visibility, verify the user-visible surface before saying the issue is fixed. A local file, commit, tag, or script output is not enough if the user's concern is what they can see in an app or public page.
For non-trivial failures, write or report:
Agent self-debug report:
- Failure:
- Root cause:
- Recovery action:
- Evidence:
- Result:
- Preventive rule: