Research Session Management

When to Use

Use this protocol to manage session boundaries and ensure seamless handoffs between sequential AI chat instances for multi-month thesis or long-form research projects.

Trigger Keywords: 'new thesis session', 'session handoff', 'continue thesis work', 'pick up where left off', 'end session'.

Routing:

• Route to side-project-planning for software repositories.
• Route to startup-analyst for venture analysis.

Prerequisites

• Primary OS: Windows host (PowerShell). Ensure file paths use Windows conventions (e.g., ~\agent-skills\...).
• Local-First Agents: 2026 workflows emphasize repository-level configuration files (e.g., .pointerrules or .claudesession) to bootstrap the agent context.
• Required Inputs: Existing INDEX.md, the latest HANDOFF_conv[N-1]_to_conv[N].md, and CLAUDE_ERROR_LOG.md (or CLAUDE_ERROR_LOG_V2.md).

Procedure

Session Startup (MANDATORY every session)

Execute in order. Do NOT skip steps. Do NOT begin work before completing all of the following steps:

1. Read the INDEX file (or INDEX_LITE if the full INDEX exceeds context limitations).
2. Read the HANDOFF file relevant to the current chapter/section. Do NOT default to the "most recent handoff" if it refers to a different chapter. If the researcher specifies 3.3.2, search for HANDOFF files referencing Chapter 3 or 3.3 before falling back to the most recent overall handoff.
3. Read all files listed in that handoff's mandatory reading list.
4. Read the Error Log (CLAUDE_ERROR_LOG.md or CLAUDE_ERROR_LOG_V2.md). This is NON-NEGOTIABLE. If no error log exists in the Knowledge Base (KB), create one immediately before starting work.
5. Read the OPS file for the current chapter (e.g., CH3_WRITING_OPS.md) if one exists. Chapter-specific error triggers and Writing Mode decisions live there.
6. Scan auto-memory for standing rules. The project's MEMORY.md index contains persistent feedback files. Before writing, identify any feedback_* entries that apply to the current task and load them. Confirm awareness of standing rules (e.g., docx run-level editing, no mid-batch method switches, mandatory visual check for visual output, define "correct" before writing fix code).
7. Initiate active research or writing only after completing steps 1 through 6.

Session Shutdown (MANDATORY every session)

Execute the following steps at the conclusion of every session:

1. Save all new notes, data, and analyses as numbered KB files on the filesystem. Do not leave key information only in the conversation history.
2. Write a HANDOFF file for the next session following the Handoff Template below.
3. Update the INDEX file by adding a conversation log entry and updating file statuses.
4. Verify that all files saved correctly by reading back a few lines from each newly written or modified file.

File Organization

• Sequential Numbering: Prefix KB files with 01_, 02_, ... NN_. Check the INDEX for the next available number.
• Subdirectories: Use subdirectories per chapter: Chapter_X/cited/, Chapter_X/nocite/, Chapter_X/archive/.
• Ownership: Markdown notes, JSON files, handoffs, and analyses are AI-maintained. Word documents are researcher-controlled; only modify them upon explicit request.
• Archiving: Move superseded versions to the archive/ directory with a version suffix before overwriting.

INDEX File Template

# Project Knowledge Base  INDEX

## 1. File Manifest
| # | File | Description | Status |
|---|------|-------------|--------|
| 01 | [name] | [description] | Complete / In progress / Needs revision |

## 2. Conversation Log
| Date | Session # | Work Done |
|------|-----------|-----------|

## 3. Next Steps
1. [priority item]

Handoff File Template

Handoff file naming convention: HANDOFF_conv[N]_to_conv[N+1].md

# Handoff: Session N  Session N+1

## What happened
[2-5 sentences summarizing progress]

## Key decisions
- D1: [decision and rationale]

## Files created or modified
- [filename]  [what changed]

## Unresolved items
- U1: [item, why it matters, suggested approach]

## Mandatory reading for next session
1. [file]  [why]

## Critical lessons
- [mistake made this session to avoid next time]

Pitfalls

• Missing Handoff: If the previous handoff is missing, scan the Git log or file modification timestamps for files changed in the last 24 hours to reconstruct the session state.
• File Collision: If two files share a sequence number, rename the older version to _v1 and move it to archive/.
• INDEX vs INDEX_LITE: Use INDEX_LITE if the file manifest contains more than 50 files or the INDEX size exceeds 15,000 tokens.
• Session End Trigger: End the session immediately if token count exceeds 50% of the model's optimal prompt caching limit, or if the reasoning latency increases significantly.
• Context Degradation: End the session and trigger shutdown when the AI shows signs of context degradation (e.g., forgetting instructions or repeating sentences).
• Handoff Accuracy: Do NOT write handoffs from memory. Re-read the actual files before summarizing. Verify all claims against raw data or final output, not earlier planning documents.
• Source Segregation: Segregate verified primary sources (citations) from speculative drafts to prevent hallucinations from polluting the core research assets.

Verification

• Zero file index numbering collisions. Check the INDEX before creating new files.
• Every handoff must cite at least one specific verified document.
• The handoff must declare exactly one starting hook for the subsequent session to prevent cold-start latency.
• File integrity: Verify that all files saved correctly by reading back a few lines from each newly written or modified file.