Agency client handoff
Package a completed engagement for transfer to the client's in-house team, DAM, or next agency — with the asset set, prompt library, model pins, rights documentation, and everything else the receiving team needs to regenerate consistent work without you in the loop.
Engagement done → reproducible handoff bundle. Knowledge transfer over lock-in: the receiving team should be able to keep the brand on-model without calling the agency back.
When to Use
- End of an engagement, client taking work in-house
- Transition to a new agency of record — outgoing handoff
- Completed pitch / campaign — asset package with provenance for the client's DAM
- Contract-mandated "source file" delivery with regeneration rights
- Year-end portfolio snapshot for a long-term retainer client
Do not use for work-in-progress reviews (use Drive links), approvals (use proof PDFs), or internal archival (use your agency's backup). Handoff is the formal, client-owned, reproducible bundle.
Prerequisites
Ask the user (one message):
- Client slug + engagement scope — which client, which project or retainer period is being handed off
- Receiving team — in-house creative? Another agency? Which format and tools do they use (Figma, Adobe, CapCut, Notion)?
- Rights status — are all inputs licensed? Any stock, fonts, or reference faces that need redacting or citing?
- Regeneration expectation — does the client need to re-generate assets on their own, or is this a static archive?
- Confidentiality — any internal prompts, competitor references, or tagged metadata that must be stripped before delivery?
- Delivery format — zip + Drive link? Direct to client S3 / DAM? Physical drive?
If the client needs regeneration capability, they'll need the gen-ai CLI themselves — add an install guide to the bundle.
How to Run
1. INVENTORY → pull every asset + results.json from clients/<slug>/ across the engagement
2. FILTER → exclude drafts, rejected variants, internal-only WIP
3. STRIP → white-label: remove agency tags, internal prompt notes, competitor refs
4. PIN MODELS → lock every prompt's model to an exact ID + version (not "latest")
5. DOCUMENT → generate README, RIGHTS.md, CHANGELOG, prompt library
6. PACKAGE → zip with consistent folder structure + naming
7. VERIFY → extract and cold-test: can a fresh machine regenerate one asset?
8. DELIVER → upload to client destination, share link, brief the receiving team
Rules:
- Pin every model ID. → (or the exact ID returned by ). "Latest" will drift and break reproducibility.
- Strip internal metadata. Tags like or are internal — not for client eyes.
- Include prompts, not just outputs. A prompt library lets the client's in-house team iterate without re-inventing the system.
- Never ship a handoff you haven't cold-tested. Unzip, pick one prompt, regenerate — if the output drifts, something isn't pinned.
Quick Reference
json
{
"handoff_kind": "client-final",
"client_slug": "acme-fintech",
"engagement": "2025-Q3 to 2026-Q2 retainer",
"delivered_at": "2026-04-22",
"brand_system": "docs/brand-system.json",
"brand_rules": "docs/brand.md",
"pinned_models": "docs/model-pins.json",
"assets_count": 247,
"prompts_count": 54,
"rights_status": "all generated, no licensed stock",
"regeneration_supported": true,
"support_window_days": 30
}
Include this as
docs/handoff-manifest.json
— a single file that tells the receiving team what they got and what they can do with it.
Quick Reference
Handoffs don't generate new assets — they package existing ones. But for reproducibility checks:
| Task | Model | Notes |
|---|
| Dry-run regeneration test (cheapest) | | Use to verify payload validity, not output quality |
| Full regeneration proof (one asset) | Same as original, pinned | Matches the original bundle; compare outputs visually |
| Re-export at higher resolution | | If the client DAM wants 4K versions |
Never substitute a newer model for the pinned one during handoff — the whole point is that the client can reproduce the agency's output exactly.
Procedure
- Reproducibility is the deliverable. If the client can't regenerate without you, the handoff failed — even if the asset folder is beautiful.
- Strip, don't redact. Remove internal tags entirely; the bundle should look native to the client.
- Version the handoff — , . Don't overwrite if an updated bundle is requested later.
- Include a sunset date for agency support — clarifies when the client self-serves or re-engages.
- Pin exact model IDs with
gen-ai models info <id> --json
- Test on a clean machine. Your laptop has tokens, fonts, caches. The bundle has to work without those.
- No lock-in flags. Everything in the bundle runs on a stock install.
- Ship prompts, brand, and rationale. Context beats artifacts for knowledge transfer.
Pitfalls
- Unpinned models — client regenerates 6 months later, output drifts, blames the agency.
- Internal metadata leakage — competitor refs, pitch-concept tags, reviewer names surviving in prompt files.
- Missing rights documentation — client legal can't sign off on re-use, bundle is DOA.
- Proprietary tools in the regen path — strip to stock CLI only.
- No cold test — bundle ships, step 1 of the README fails, urgent Slack follows.
- One-and-done mentality — promise 30 days of Q&A post-handoff, price it in.
Verification
Run
to confirm authentication, then re-run the failed command with
.
Step 1: Inventory and filter
bash
CLIENT="acme-fintech"
HANDOFF="handoff/$CLIENT-$(date +%Y-%m-%d)"
mkdir -p "$HANDOFF"/{assets,prompts,source,docs}
find "clients/$CLIENT" -name "results.json" -exec cp {} "$HANDOFF/source/" \;
cp -r "clients/$CLIENT/deliverables/"* "$HANDOFF/assets/"
node scripts/extract-prompts.js "$HANDOFF/source/" > "$HANDOFF/prompts/prompt-library.json"
Step 2: Pin models, strip internal data
bash
# Get pinned model IDs for every model the engagement used
for model in $(jq -r '.jobs[].model' "$HANDOFF/source/"*.json | sort -u); do
gen-ai models info "$model" --json >> "$HANDOFF/docs/model-pins.json"
done
# Strip internal tags from the prompt library
jq 'del(.jobs[].tags[] | select(startswith("internal-") or startswith("retainer-")))' \
"$HANDOFF/prompts/prompt-library.json" > "$HANDOFF/prompts/prompt-library.clean.json"
mv "$HANDOFF/prompts/prompt-library.clean.json" "$HANDOFF/prompts/prompt-library.json"
Step 3: Documentation
Generate a
covering: folder walkthrough, regeneration steps (CLI install + one worked example), pinned model note, rights status, support contact + sunset date. Include
,
,
,
,
.
Minimum regeneration steps to include in the README:
1. Install CLI: curl -fsSL https://picsart.com/gen-ai-cli/install.sh | bash
2. gen-ai login
3. Pick a prompt from prompts/prompt-library.json
4. gen-ai generate -m <pinned-model-id> -p "<prompt>"
Step 4: Package + verify
bash
cd handoff
zip -r "$CLIENT-handoff-$(date +%Y-%m-%d).zip" "$CLIENT-$(date +%Y-%m-%d)/" \
-x "*.DS_Store" "*/.git/*"
# Cold test — extract to a clean dir, regenerate one asset
tmpdir=$(mktemp -d) && unzip -q "$CLIENT-handoff-$(date +%Y-%m-%d).zip" -d "$tmpdir"
cd "$tmpdir/$CLIENT-$(date +%Y-%m-%d)"
SAMPLE_PROMPT=$(jq -r '.jobs[0].prompt' prompts/prompt-library.json)
SAMPLE_MODEL=$(jq -r '.jobs[0].model' prompts/prompt-library.json)
gen-ai generate -m "$SAMPLE_MODEL" -p "$SAMPLE_PROMPT" --dry-run --debug
# Dry-run validates = bundle is regeneration-ready.
Cost & time
| Phase | Spend | Time |
|---|
| Inventory + filter + strip | $0 | 1-2 hrs |
| Documentation (README, RIGHTS, CHANGELOG) | $0 | 1-2 hrs |
| Cold-test regeneration (1-2 assets) | ~$0.50 | 15 min |
| Total handoff | <$1 | ~4 hrs |
The real cost is time, not credits. Budget a half-day per handoff; cutting corners creates support calls for months after.
See also
workflows/agency-brand-scoping/
workflows/agency-multi-brand-pack/
workflows/agency-pitch-mockups/
- — regeneration reference for the client