Creating Handoffs

Creating Handoffs

Produce a handoff document that transfers full working context to the next session with no information loss.

Interaction mode

This skill leans Direct by default. For the full Collaborative-vs-Direct protocol and override rules, see the Interaction Modes reference in the ai-research-workflows:using-research-workflows skill.

Process

1. Gather context

Run these in parallel where possible:

Git state:

• Current branch name
• Current commit hash (short)
• Summary of uncommitted changes (git status and git diff --stat)

Workflow artifacts — search docs/rse/specs/ (then legacy .agents/) for:

• research-*.md — research documents
• plan-*.md — plan documents
• experiment-*.md — experiment documents
• implement-*.md — implementation documents
• validation-*.md — validation reports
• handoff-*.md — previous handoff documents

Session context:

• Review the conversation to understand what tasks were worked on.
• Identify the current workflow phase (Research, Plan, Iterate Plan, Experiment, Implement, Validate).
• Note which workflow artifacts were produced or referenced this session.

Research state (for research/scientific work):

• Random seeds, the active environment/lockfile (pixi.lock / uv.lock), and dataset versions/paths/checksums in play.
• Partial or intermediate results, saved checkpoints/outputs, and any long-running experiment or job still in flight — with how to check or resume it.

2. Determine what's relevant

From the gathered context, identify:

• Tasks: What was being worked on and the status of each (completed, in progress, planned).
• Current phase: Where in the workflow cycle the work sits.
• Artifacts: Which docs/rse/specs/ documents are relevant.
• Critical files: The 2–3 most important files the next session must read first.
• Recent changes: What code was modified (use file:line references).
• Learnings: Important discoveries, patterns, or gotchas.
• Research state: seeds, environment/lockfile, data versions, and partial results needed to resume (see ai-research-workflows:ensuring-reproducibility).
• Verification status: what is tested and passing vs. broken or unverified, and what is uncommitted or unpushed.
• Next steps: What the next session should do, in priority order.

Report true state — don't paper over broken work

A handoff must reflect reality, not a tidy fiction. Before generating the document, explicitly surface:

• Uncommitted / unpushed work — name it; the next session cannot see your working tree.
• Failing or unrun tests — state which, with the actual status; never imply green when it isn't.
• Unreproducible or unverified results — flag any result that has not been reproduced or whose provenance is incomplete.

Put these in a "Known-broken / unverified" callout in the handoff. A clean-looking handoff that hides broken state is worse than none — it sends the next session down a false trail.

3. Generate the handoff document

Filename format: handoff-YYYY-MM-DD-HH-MM-<slug>.md

• YYYY-MM-DD-HH-MM is the current date and time.
• <slug> is a brief kebab-case description of the work.
• Example: handoff-2025-06-15-14-30-auth-system-refactor.md

Read the handoff template: assets/handoff-template.md

Fill out all sections:

• Replace all placeholder text with actual content.
• Remove artifact sections that don't apply (e.g., if no experiments were run, remove Experiment Reports).
• Be specific — use file:line references, not vague descriptions.
• Name the recommended next skill for the receiving session based on the current phase (e.g., ai-research-workflows:implementing-plans, ai-research-workflows:validating-implementations).

Save to docs/rse/specs/handoff-YYYY-MM-DD-HH-MM-<slug>.md.

4. Present the handoff

After saving, present a concise summary:

## Handoff Created

**File:** `docs/rse/specs/handoff-<filename>.md`
**Current Phase:** [phase]
**Status:** [brief status of work]

### Quick Summary
[2–3 sentence summary of what was done and what's next]

### For the Next Session
Start by running:
> Read the handoff document at `docs/rse/specs/handoff-<filename>.md` and resume the work described within.

Or to continue with the workflow:
> [recommended-skill] [relevant arguments]

Writing guidelines

• More information, not less. The template is the minimum; always include more when necessary.
• Be thorough and precise. Include both top-level objectives and lower-level details.
• Avoid excessive code snippets. Prefer path/to/file.ext:line references. Only include code blocks when describing an error being debugged or a critical pattern.
• Cross-reference workflow artifacts. Link to research, plan, experiment, and implementation documents by filename so the next session can read them.
• Name the recommended next skill. Based on where you are in the workflow, tell the next session which skill to invoke next (e.g., ai-research-workflows:implementing-plans, ai-research-workflows:validating-implementations).
• Include learnings. Capture non-obvious insights about the codebase, patterns that matter, or gotchas encountered.

Common Mistakes

• Dumping raw history instead of a focused summary — pasting the full conversation or every git commit message produces an unreadable wall of text. Distil to what the next session actually needs to act.
• Omitting the critical files to read first — the most common cause of a slow hand-off restart is not knowing where to begin. Always name the 2–3 files the next session must read before doing anything else.
• Not naming the recommended next skill — leaving "next steps" vague forces the receiving session to re-derive the workflow position. Always state the exact skill to invoke next (e.g., ai-research-workflows:implementing-plans).
• Missing learnings and gotchas — documenting what was done without capturing why certain decisions were made, or what dead ends were hit, forces the next session to rediscover them.
• A clean-looking handoff that hides broken state — omitting failing tests, uncommitted work, or unverified results sends the next session down a false trail; surface them in a "Known-broken / unverified" callout.
• Omitting research state — leaving out seeds, environment/lockfile, dataset versions, partial results, or in-flight jobs makes the work impossible to resume or reproduce; capture them for any research/scientific work.

Cross-references

• Use ai-research-workflows:implementing-plans as the recommended next skill when a plan is approved and ready to execute.
• Use ai-research-workflows:validating-implementations when implementation is complete and needs verification.
• Use ai-research-workflows:planning-implementations or ai-research-workflows:iterating-plans when the next session needs to (re-)design the approach.
• All handoff documents are picked up automatically by ai-research-workflows:using-research-workflows when it surveys docs/rse/specs/ (and legacy .agents/).