/handoff is my new favourite skill

Actionable Insights

• Create a /handoff workflow for context slicing, not just summarization. The skill should write a compact markdown artifact for a fresh agent with goal, current state, relevant files, constraints, decisions, failed attempts, next actions, and explicit non-goals. Use Matt Pocock’s source as a reference via AI Hero skills and mattpocock/skills. First step: add docs/handoffs/ or a temp handoff folder and require filenames like handoff-YYYYMMDD-feature-purpose.md. Evaluate by whether a fresh session can continue without reading the full transcript.

• Use handoff when a side quest appears during planning. Around 5:37–7:09, the creator describes extracting an out-of-scope bug or issue from a grilling session into a separate agent/session. Practical checklist: name the reason for the handoff, define what should be included, specify the target output (GitHub issue, prototype, PR, research memo), and keep the parent session focused. Caution: if the side quest changes the main goal, return the result to the parent before continuing.

• Pair /handoff with /prototype for uncertain UI or logic. The video shows a prototype branch/session reaching 169K tokens, then sending learnings back to the original grilling session. First experiment: during requirements discovery, hand off only the uncertain interaction to a prototype agent and request screenshots, test notes, and lessons learned. Evaluate by whether the parent plan improves without bloating the original context.

• Prefer handoff over auto-compact when you need selective context. The creator’s distinction is useful: compact preserves a long single session; handoff chooses the slice another session needs. Use compact for continuity; use handoff for branching, parallel exploration, or passing work to another harness. Caution: do not duplicate entire transcripts; include pointers to existing artifacts instead.

• Standardize resume prompts for your own machine. Commenters add a strong practitioner pattern: save handoffs to known locations and use aliases to start a new agent with the handoff loaded automatically. Example shape: agent --resume docs/handoffs/latest.md or a shell alias for your harness. Evaluate by time-to-restart after context loss and number of clarification questions the new agent asks.

Core thesis

The video argues that a handoff skill is valuable because modern agent work is not one linear conversation. You often need to branch work into separate sessions, preserve the parent context, prototype uncertain areas, file issues, or resume later. A good handoff document deliberately compresses only the relevant slice of context so another agent can act without inheriting the full noisy conversation.

Big ideas / key insights

• Context windows have practical “smart” and “dumb” zones. The creator claims performance degrades before the official maximum context limit, so context budgeting matters.
• Compact and handoff solve different problems. Compact continues a session; handoff transfers a selected slice to another session.
• Handoff sharpens the current conversation. By declaring what is out-of-scope and passing it elsewhere, the parent session stays focused.
• DIY subagents can be built from handoff loops. Parent session → handoff to child → child returns handoff with learnings.
• The skill is simple but valuable because it makes a repeated habit reliable. Several commenters already do this manually, which supports the pattern’s usefulness.

Best timestamped moments with interpretation

• 0:31 — The handoff skill is introduced as a markdown summary for a fresh agent. Interpretation: the artifact is intentionally simple.
• 1:31–3:34 — Context window and compaction explanation. Interpretation: the “smart zone” framing is anecdotal but practically useful.
• 4:36–5:37 — The problem: wanting to branch a side task without polluting the current session. This is the core use case.
• 6:39 — Defining why the handoff exists helps the current grilling session collapse scope. This is a subtle but important benefit.
• 7:09–8:45 — Handoff to prototype and then handoff back. Interpretation: a practical multi-session pattern that approximates subagent orchestration.
• 10:17 — Avoid duplicating content already captured elsewhere. Interpretation: good context hygiene.

Practical takeaways / recommended workflow

1. Add a handoff template with: objective, background, current state, relevant files, constraints, open questions, next command, and done criteria.
2. During planning/grilling, mark anything out of scope as handoff candidate instead of discussing it immediately.
3. Spawn a separate session for each candidate only when it has a clear target artifact.
4. Ask the child session to write a return handoff summarizing learnings and recommended parent-session changes.
5. Archive handoffs near the project only if they have lasting value; otherwise use a temp directory.
6. Keep handoffs short and link to docs/PRs/issues rather than copying everything.

Comment insights

The comments are unusually useful. The highest-liked comment says many users already ask agents to save handoff docs manually and jokes that one may not need a /handoff skill. Matt Pocock replies that he did that for a while but having a skill is nicer, which supports the “codify repeated behavior” thesis. Other commenters mention OpenCode conversation forks, known-location handoff folders, quick handoff folders for subagents, and “resume prompt” variants. The pushback is that this may be common knowledge rebranded; the practical answer is that repeatability and integration are the value, not novelty.

Deep research on the main claims

• Claim: Matt Pocock’s skills include /handoff and related workflow skills. Support: AI Hero’s skills page lists /handoff, /prototype, /review, /tdd, /grill-with-docs, /to-prd, and /to-issues, and points to mattpocock/skills. Contradiction/caution: the video is itself a creator explanation, and independent outcome data is limited.

• Claim: long context can degrade practical agent quality before maximum context. Support: this aligns with common LLM engineering experience around attention dilution and context management, and Anthropic/agent best-practice discussions often recommend compacting or managing context. Contradiction/caution: exact thresholds such as 120K tokens are subjective and model/harness dependent.

• Claim: handoff differs from compaction. Support: the transcript gives a clear functional distinction, and many harnesses’ compact/resume features are designed for continuing the same task rather than branching selective subwork. Contradiction/caution: some tools now offer forks, subagents, or branchable sessions, which can reduce the need for a custom handoff skill.

• Claim: handoff can emulate subagents. Support: the transcript’s parent-child-parent prototype pattern and comments about handoff folders/subagents support this. Contradiction/caution: real subagent runtimes can manage isolation, tool permissions, status, and merge-back more robustly than ad-hoc markdown.

Verdict

• "/handoff is worth making a skill." — Agree, medium-high confidence. If you do it repeatedly, a skill improves consistency. If you rarely branch sessions, a plain prompt may be enough.
• “Context has a smart zone and dumb zone.” — Mixed, medium confidence. The practical warning is valid; the precise threshold is anecdotal.
• “Handoff is better than compact.” — Disagree as a universal claim, high confidence. They solve different problems. Use compact for one long thread; use handoff for branching and selective transfer.
• “This is basically DIY subagents.” — Agree with caveats, medium confidence. It captures the workflow shape but lacks runtime guarantees.
• “The pattern is new.” — Disagree, high confidence. Comments show many practitioners use resume prompts/handoff docs already. The novelty is packaging and teaching it clearly.

Screen-level insights

• 0:00 — Intro/talking-head framing about packaging instincts into skills. Visual value is low but sets the creator context.
• 0:31 — Screen shows the handoff skill location under productivity and the instruction to write a handoff document. This anchors the actual artifact.
• 1:31 — Context-window explanation segment. Visual matters if diagrams are shown, but the transcript carries the technical claim.
• 6:07–6:39 — Grilling session example where a handoff is defined with reason and scope. This shows handoff as an active planning tool, not a post-hoc summary.
• 7:09–7:41 — Handoff to prototype during grilling. Visual value: demonstrates selecting a hard-to-answer question for code exploration.
• 8:13 — Prototype session with high token usage. This supports the point that child sessions can absorb expensive exploration.
• 10:17 — Rules about not duplicating existing artifacts. This matters because handoff quality depends on compression and pointers.

My read / why it matters

This is a small pattern with big leverage. It prevents the two common agent-session failure modes: letting side quests bloat the main context, and losing useful state when starting fresh. The best implementation is not a long template; it is a short, disciplined handoff with explicit target output and acceptance criteria.

Verification notes

Four checks were applied. Source/evidence audit: Matt Pocock/AI Hero skill links were verified and context-window claims were caveated. Transcript/comment/frame fidelity audit: all workflow examples come from extracted timestamps and comments. Hallucination/overclaim audit: novelty and threshold claims were softened where unsupported. Actionable Insights audit: bullets include implementation fields, file paths, evaluation criteria, and cautions. Residual uncertainty: exact harness commands for compact/fork vary, so adapt the template to your tool.