indranilbanerjee/engagement-workflow

/digital-marketing-pro:engagement-workflow — 12-Part Engagement Orchestrator

This skill orchestrates the full marketing engagement using the 12-Part sequential methodology. Every brand engagement runs through the same 12 parts in sequence, producing a canonical set of files at each stage.

Context efficiency

Heavy skill. Grep before Read any referenced file, then Read only matched ranges with offset + limit. List the brand's workspace at ~/.claude-marketing/brands/{slug}/ (or $CLAUDE_PLUGIN_DATA/digital-marketing-pro/brands/{slug}/ when that env var is set) before opening files. On re-invocation mid-session, skip files already in context.

Read these references before producing output:

• engagement-flow-methodology.md — the full 12-Part flow
• two-views-model.md — v1 / v2 architecture
• stone-vs-opinion.md — confidence tagging
• decision-matrix-rerun.md — when to re-run what
• update-back-rule.md — versioning protocol
• living-instruction-file-spec.md — LIF schema

Operating Mode

This skill is invoked via the /digital-marketing-pro:engagement command family. The command is a thin router — this skill is the single source of truth for the engagement lifecycle, the checkpoint protocol, and the per-part production contract. Each subcommand maps to a specific lifecycle action. The skill calls engagement-state.py for persistence via:

python "${CLAUDE_PLUGIN_ROOT}/scripts/engagement-state.py" <subcommand> ...

You should never hand-edit _engagement.json — always go through engagement-state.py.

Checkpointing & Resume (single source of truth)

Every long engagement run is resumable. The checkpoint protocol is: init a run → save each part as it completes → finalize → publish to the visible output folder. This lets an interrupted run (context exhaustion, user cancel, machine sleep) resume from the next un-checkpointed part instead of restarting from Part 1.

1. On start, after the brand pre-condition passes, open a checkpoint run and link it to engagement state:

python "${CLAUDE_PLUGIN_ROOT}/scripts/checkpoint-manager.py" init \
    --brand "{brand_slug}" --workflow engagement --topic "{engagement_id}"

# Record the returned run_id into _engagement.json so resume can find it:
python "${CLAUDE_PLUGIN_ROOT}/scripts/engagement-state.py" set-checkpoint-run \
    --brand "{brand_slug}" --id "{engagement_id}" --run-id "{run_id}"

set-checkpoint-run stores the run_id in _engagement.json, making the resume linkage real (previously the run_id was never persisted).

2. After each part completes and passes its quality gate, the orchestrator saves that part's output:

python "${CLAUDE_PLUGIN_ROOT}/scripts/checkpoint-manager.py" save \
    --brand "{brand}" --run-id "{run_id}" \
    --step {part_number} --content-file "{path_to_that_part_deliverable}" --extension md

Pass the actual deliverable path for that part (e.g. Part 3 saves the Four Core Documents path; Part 8 saves the Growth Plan path) — never a placeholder for a different part.

3. Before saving Part 5 (Client Validation) and Part 8 (Growth Plan) deliverables, run the full quality gate:

# BLOCKING gate — Part 5 and Part 8 deliverables cannot be checkpointed until this passes
/digital-marketing-pro:check "{path_to_deliverable}" --full --brand {brand}

If /digital-marketing-pro:check --full returns BLOCKED, fix the CRITICAL issues before checkpointing the part.

4. After the final part, publish every artifact to the user-visible folder and finalize:

python "${CLAUDE_PLUGIN_ROOT}/scripts/output-publisher.py" publish-run \
    --brand "{brand}" --run-id "{run_id}"

python "${CLAUDE_PLUGIN_ROOT}/scripts/checkpoint-manager.py" finalize \
    --brand "{brand}" --run-id "{run_id}" --status completed

Then point the user at the visible output folder via /digital-marketing-pro:output-folder {brand}.

To resume an interrupted run, use /digital-marketing-pro:resume — it reloads every saved part and continues from the next un-checkpointed part.

State validation & rework caps

• Validate a part's outputs against the manifest before marking it complete:

  python "${CLAUDE_PLUGIN_ROOT}/scripts/engagement-state.py" validate-part \
      --brand "{brand}" --id "{id}" --part {N}
  

  This diffs the actual files on disk against the PART_DEFINITIONS manifest and flags missing deliverables. Use it in file-tree and before next.

• Repair a partially-initialised engagement directory (instead of crashing on a non-empty dir):

  python "${CLAUDE_PLUGIN_ROOT}/scripts/engagement-state.py" init --repair \
      --brand "{brand}" --id "{id}"
  

  --repair completes the canonical directory tree and state file on a dir that holds only partial state.

• v2 re-run cap: a maximum of 2 v2 re-run rounds per part is allowed without explicit user override. The round count is stored in _engagement.json. If a part would exceed 2 rounds, stop and ask the user to explicitly approve further re-runs (records the override in state). This prevents unbounded re-run loops.

Subcommands

/digital-marketing-pro:engagement start <brand-slug> <engagement-id>

Purpose: Initialise a new engagement.

Steps:

1. Validate that the brand profile exists at ~/.claude-marketing/brands/{brand-slug}/profile.json. If not, instruct the user to run /digital-marketing-pro:brand-setup first.
2. Run python ${CLAUDE_PLUGIN_ROOT}/scripts/engagement-state.py init --brand {brand-slug} --id {engagement-id}.
3. Confirm the directory tree was created and report the next required action (Part 1 intake).
4. Walk the user through Part 1 Stone vs Opinion intake by asking the questions one batch at a time.

Part 1 intake questions (ask in this order):

Stone — what the client knows for certain:

1. Company basics: founded year, employee count, headquarters location, geographic operations
2. Business model: revenue streams, pricing tiers, primary product/service categories
3. Current marketing: channels currently active, monthly marketing spend, current measurable KPIs
4. Tech stack: CRM, email service provider, analytics setup, ad accounts
5. Customer base scale: customer count, biggest named customer, average order value if known

For each Stone fact, capture:

• The fact itself
• Source (how the user knows / what document confirmed it)

Save each via:

python ${CLAUDE_PLUGIN_ROOT}/scripts/engagement-state.py add-stone-fact --brand {slug} --id {id} --fact-json '{"category":"...","fact":"...","source":"..."}'

Opinion — what the client believes:

1. Brand positioning: how does the client describe their position in the market?
2. Customer base: who do they think their customers are? Why do they buy?
3. Competitors: who do they consider their main competitors?
4. Growth opportunities: where do they think the biggest opportunity is?
5. What is working: what marketing activity does the client believe is working?
6. What is not working: what does the client believe is not working?

For each Opinion, capture:

• The hypothesis
• Client's evidence for it (could be intuition, anecdote, partial data)
• Research question — what would the unbiased research need to verify or refute?

Save each via:

python ${CLAUDE_PLUGIN_ROOT}/scripts/engagement-state.py add-opinion --brand {slug} --id {id} --hypothesis-json '{"category":"...","hypothesis":"...","client_evidence":"...","research_question":"..."}'

On completion of Part 1: mark Part 1 as completed via mark-part-completed --part 1, advise the user to proceed to Part 2 (External Research).

/digital-marketing-pro:engagement next [brand] [id]

Purpose: Advance to the next part.

Steps:

1. Read engagement status via engagement-state.py status
2. Identify the current part and next not-yet-completed part
3. Confirm with the user that the current part is genuinely complete (do not auto-advance — ask)
4. On confirmation, mark current as completed, advance current_part pointer
5. Brief the user on what the new part requires

/digital-marketing-pro:engagement status [brand] [id]

Purpose: Show engagement status.

Steps:

1. Run engagement-state.py status — get the full state
2. Read the Living Project Instruction File header
3. Format a human-readable summary:
   • Engagement: brand + id + start date
   • Current part: part name + days in
   • Completed parts: list
   • Pending parts: list
   • Open re-run decisions: count
   • LIF last updated: date
4. If the engagement has open items needing resolution, list them

/digital-marketing-pro:engagement file-tree [brand] [id]

Purpose: Show the engagement directory file tree.

Steps:

1. Run engagement-state.py file-tree
2. Format as an indented tree
3. Highlight files that are missing per the canonical structure. Use engagement-state.py validate-part --part {N} to diff each completed part's actual files against the PART_DEFINITIONS manifest (e.g., if Part 3 is marked completed but 3.1-business-and-sbu-analysis.md is missing, validate-part flags it deterministically instead of eyeballing).

/digital-marketing-pro:engagement validate [brand] [id]

Purpose: Run the Part 5 Client Validation flow.

Pre-condition: Parts 2, 3, 4 must be completed.

Steps:

1. Verify pre-conditions (Parts 2, 3, 4 completed)
2. Invoke the client-validation-document skill — it produces the Part 5 deliverable: a structured document presenting each finding from v1 with ACCEPT/REJECT/EDIT/DEFER options
3. Run the full quality gate on the Part 5 deliverable before it goes to the client: /digital-marketing-pro:check "{part5_path}" --full --brand {brand}. If it returns BLOCKED, fix the CRITICAL issues first (this gate is mandatory before Part 5 and Part 8 deliverables).
4. After the user reviews and provides decisions, parse them into a triggers list per the Decision Matrix categories
5. Run engagement-state.py decision-matrix --triggers "{comma-separated}" to compute the v2 re-run plan
6. Present the re-run plan to the user
7. Mark Part 5 completed; on user approval of the re-run plan, advance to Part 6

/digital-marketing-pro:engagement re-run-decision [brand] [id]

Purpose: Apply the Decision Matrix to compute v2 re-runs.

Steps:

1. Read the Part 5 Client Validation Document
2. Categorise rejected/edited findings into Decision Matrix triggers
3. Show the triggers and the computed re-runs
4. Estimate the cost (rough token count) of each re-run
5. Await user approval — they can accept, modify (skip some, add others), or reject
6. Record the executed plan via engagement-state.py record-rerun-execution

/digital-marketing-pro:engagement update-back [brand] [id] --doc <doc-id> --reason <reason>

Purpose: Apply the Update-Back Rule to bump a source document version after Part 7+.

Pre-condition: The user has already drafted the corrected document content.

Steps:

1. Read the current version of the doc
2. Confirm the correction with the user (validation step per the Update-Back Rule)
3. Bump the version via engagement-state.py bump-version --doc {id} --reason "{reason}"
4. Save the new version file with a header noting v(prev) → v(new) changes
5. Update the Living Project Instruction File via lif-log-change — it now appends the change to living-instruction-file.md and refreshes the header date, so the LIF reflects the correction immediately
6. Identify downstream documents that may need review and add to the engagement's review queue

/digital-marketing-pro:engagement lif-show [brand] [id]

Purpose: Display the Living Project Instruction File.

Steps: Run engagement-state.py lif-show and format the markdown output for readability.

/digital-marketing-pro:engagement list-engagements [brand]

Purpose: List all engagements (optionally filtered by brand).

Steps: Run engagement-state.py list-engagements --brand {slug} and format as a table.

Production shorthands

The command family also exposes four production shorthands that route straight to the part-producing skills (documented in Per-Part Production Targets below). These match the command surface one-to-one:

• /digital-marketing-pro:engagement four-core <brand> <id> [--doc 3.X] [--view v2] [--combined] — Part 3, invokes the four-core-documents skill
• /digital-marketing-pro:engagement growth-plan <brand> <id> — Part 8, invokes the growth-plan skill
• /digital-marketing-pro:engagement yearly-planner <brand> <id> — Part 8 companion, invokes the yearly-planner skill
• /digital-marketing-pro:engagement loop <brand> <id> — Part 12, invokes the continuous-improvement-loop skill

Per-Part Production Targets

Each part is produced by real, existing agents and skills. This orchestrator dispatches to the targets below — there are no wrapper skills named external-research / preparation-documents / channel-strategy-fanout / execution-artefacts / ai-creative-instructions; those never existed. Use the exact targets named here:

| Part | Real target(s) | |------|----------------| | 1 | (this skill — intake walked here directly) | | 2 | agents market-intelligence + competitive-intel; skill audience-intelligence (invoke as a skill); reference compliance-rules.md (load as context) | | 3 | skill four-core-documents (produces 3.1, 3.2, 3.3, 3.4) | | 4 | command competitor-analysis + skills audience-intelligence + agent market-intelligence | | 5 | skill client-validation-document | | 6 | re-runs invoke skill four-core-documents with --view v2 | | 7 | skills content-engine + campaign-orchestrator + analytics-insights | | 8 | skills growth-plan + yearly-planner | | 9 | per-channel skills — paid-advertising, aeo-geo, social-strategy, seo-plan, email-sequence (one per channel family) | | 10 | skill content-engine (execution / output mode) | | 11 | skills content-engine + ad-creative + video-script (creative briefs); asset rendering happens in your own creative tooling (design team, AI image/video tools, or a connected design platform), then finished assets are signed via c2pa-metadata | | 12 | skill continuous-improvement-loop |

Parallel Dispatch

Several parts of the engagement contain independent sub-tasks that should be dispatched in parallel via multiple Task tool calls in a single message — not sequentially. Dispatching independent sub-tasks concurrently is substantially faster than running them one after another; actual time varies by engagement depth, model, and rate limits. Keep concurrent subagents to a handful (roughly 3–8) — past that you queue against API rate limits and the win drops; under 3 there is nothing to parallelize.

Cost note: total token usage is broadly similar (you're doing the same work) but billed-per-turn input costs trend up slightly because each parallel subagent re-loads its context.

Parts that benefit from parallel dispatch:

| Part | Parallel-eligible work | How to dispatch | |---|---|---| | Part 2 — External Research | Market sizing, competitor landscape, customer signals, regulatory landscape — none depend on each other | Dispatch the market-intelligence agent and the competitive-intel agent as parallel Task calls; invoke audience-intelligence as a skill; load compliance-rules.md (a reference file) as context — not as a subagent | | Part 4 — Competitive + Customer + Market | Four documents (4.1, 4.2, 4.3, 4.4) are independent — they reference Part 2 only | Dispatch all four in a single message with the four respective subagents | | Part 9 — Channel Strategy Fan-out | Up to 17 channel docs in 7 families. Families 2 (Paid platforms), 3 (Organic & Influencer), 4 (Marketplace & CRM), 5 (Content/ATL/BTL/PR) are independent after Families 1 (Search & Campaign) and 6 (Web + Measurement) complete | Sequence: F1 → (F2 ∥ F3 ∥ F4 ∥ F5 in parallel) → F6 → F7. The middle batch is four parallel Task calls in one message. | | Part 10 — Execution Artefacts | Ad copy, post copy, headlines, CTAs across channels — independent per channel | Dispatch one subagent per channel in parallel | | Part 11 — AI Creative Instructions | Visual asset briefs — independent per asset | Dispatch in parallel per asset |

Parts that MUST stay sequential (have hard data dependencies):

• Part 1 → Part 2 (intake feeds research scope)
• Part 3 → Part 4 (Four Core Documents feed competitive/customer/market analysis)
• Part 5 → Part 6 (Client Validation drives which docs need v2 re-runs)
• Part 7 → Part 8 (prep docs feed the Growth Plan)
• Part 8 → Part 9 (Growth Plan drives channel fan-out)

Cross-cutting rules:

1. Never dispatch parallel agents that need to write to the same file simultaneously — chunk by output file.
2. Each parallel subagent gets the engagement slug and the LIF path so it can read shared context, but writes ONLY to its own numbered per-part subdirectory (01-… 12-…).
3. Subagents never mutate engagement state. A subagent must NOT call lif-log-change, mark-part-completed, bump-version, or any other engagement-state.py write, and must NOT touch _engagement.json or living-instruction-file.md. Those are unlocked read-modify-write files; concurrent writers lose updates. Each subagent returns its output as per-part files only. After a parallel batch completes, the orchestrator alone applies state mutations — one lif-log-change per batch, plus mark-part-completed / bump-version as needed — and then re-reads the LIF before the next step.
4. If a parallel batch fails partway, the failed subagent's outputs are NOT auto-rolled-back — re-dispatch only the failed ones; the successful peers stay valid.

For multi-dimensional commands outside the 12-part flow (e.g. /digital-marketing-pro:competitor-analysis, /digital-marketing-pro:seo-audit, /digital-marketing-pro:content-engine), the same pattern applies — dispatch independent dimensions in parallel via multiple Task calls in a single message.

Running an engagement in a single conversation

A large-context model can hold much of an engagement — intake, external research, the Four Core Documents (61 steps), competitive/customer/market analysis, Client Validation, selective v2 re-runs, preparation docs, Growth Plan + Yearly Planner, channel fan-out, execution artefacts, creative briefs, and the continuous-improvement loop — within one working session (a full engagement typically produces 50–60 canonical documents).

The checkpoint + persistence pattern is still the default — always. Even when everything fits in one conversation:

• engagement-state.py + checkpoint-manager.py remain the system of record: audit trail, cross-conversation resume, and multi-user / multi-day continuity all depend on persisted state.
• Do NOT skip LIF updates or state writes on the assumption that "it's all in context." An interruption still loses in-memory work, and a teammate resuming the engagement reads persisted state — not your conversation.
• The only single-conversation convenience is that you re-read fewer files mid-session because they are already in context. It does not remove the need to persist, checkpoint, and update the LIF.

Quality Discipline

1. Never hand-edit _engagement.json. Always go through engagement-state.py.
2. Never delete v1. When v2 is produced, both stay.
3. Always update the LIF when source docs change. Use lif-log-change — it appends to the change-log section of living-instruction-file.md and refreshes the header date, so the LIF never goes stale.
4. Always cite source per fact. Stone facts cite the validation source; Opinion hypotheses cite the client's evidence.
5. Never auto-advance parts. The user confirms part completion explicitly.
6. Always parallelize independent work. When a Part has 2+ independent sub-tasks (see Parallel Dispatch above), dispatch them in a single message with multiple Task calls. Sequential dispatch of independent work wastes wall-clock time and API turns.
7. Never auto-execute v2 re-runs without user approval. Show the plan, get approval, then run. Cap: 2 v2 re-run rounds per part without an explicit user override (stored in state).

Examples

Example: Starting a new engagement

User: I want to start a new engagement for Acme Corp this quarter.

Skill response:
1. Verify Acme has a brand profile (slug: acme-corp). Found.
2. Initialise engagement: engagement-state.py init --brand acme-corp --id 2026-q2
3. Engagement directory created at ~/.claude-marketing/brands/acme-corp/engagements/2026-q2/
4. Walk Part 1 intake. Begin with Stone questions:
   "Let's start with what we know for certain. What year was Acme founded? Where is it headquartered? How many employees today?"
5. (User answers; each fact captured via add-stone-fact)
6. Move to Opinion questions:
   "Now let's capture what you believe but haven't formally validated. How would you describe Acme's positioning in the market today?"
7. (User answers; each hypothesis captured via add-opinion)
8. Mark Part 1 completed. Advise: "Part 1 is complete. The unbiased External Research (Part 2) is next — recommend running it now or scheduling it for the next session."

Example: Mid-engagement status check

User: Where are we on the Beta Industries engagement?

Skill response:
- Reads engagement-state.py status --brand beta-industries --id 2026-h1
- Shows:
  Engagement: beta-industries / 2026-h1
  Started: 2026-04-01 (32 days in)
  Current part: Part 9 — Channel Strategy Fan-out (in progress)
  Completed parts: 1, 2, 3, 4, 5, 6, 7, 8
  Pending parts: 9 (in progress), 10, 11, 12
  Open re-run decisions: 1 (Part 6 v2 re-runs in progress)
  Open items requiring resolution: 2 (LinkedIn Ads CPL forecast pending; Q3 budget envelope pending)
  LIF last updated: 2 days ago
  Next required action: complete remaining Part 9 channel docs (5 of 9 complete)

Example: Applying the Decision Matrix

User: We just finished Part 5 client validation. The client rejected our positioning recommendation and added two new competitors. What re-runs do we need?

Skill response:
1. Categorise the changes:
   - Two new competitors → trigger "competitors_changed"
   - Positioning rejected → trigger "positioning_changed"
2. Run engagement-state.py decision-matrix --triggers "competitors_changed,positioning_changed"
3. Output the triggered re-runs:
   Triggered re-runs: 3.1, 3.2, 3.3, 3.4, 4.1, 4.2
   (competitors_changed triggers all four Core Docs + 4.1 + 4.2; positioning_changed adds 3.3 which is already in scope)
4. Estimate token cost: ~85K tokens for the full re-run set
5. Ask: "Approve all? Modify? Skip any?"
6. On user approval, mark Part 6 started and invoke four-core-documents skill with view=v2 for each triggered doc

Example: Update-Back correction in flight

User: We've been live for two months. Q1 channel data shows our actual CAC for Segment X is INR 4,800 — but Core Doc 3.1 v2.0 had it estimated at INR 3,000. Need to update.

Skill response:
1. Validate the correction: 60+ days of data, statistically meaningful sample. ✓
2. Read current 3.1 v2.0
3. Bump version: engagement-state.py bump-version --doc 3.1 --reason "Segment X CAC corrected from INR 3,000 to INR 4,800 based on Q1 channel data"
4. Save 3.1 v2.1 with header note: "v2.0 → v2.1 changes: Section 4 (Unit Economics) — Segment X CAC corrected. LTV:CAC drops from 4.0 to 2.5; below 3.0 health threshold."
5. lif-log-change --section "Recent Corrections" --summary "Segment X CAC corrected to INR 4,800. Channel mix for Segment X under review."
6. Identify downstream review needed: Core Doc 3.4 (DMFlow) channel allocation for Segment X may need revision
7. Add to engagement review queue

Related skills

• four-core-documents — produces Part 3 deliverables
• client-validation-document — produces Part 5 deliverable
• growth-plan + yearly-planner — produce Part 8 deliverables
• continuous-improvement-loop — handles Part 12

Related references

• engagement-flow-methodology.md
• four-core-documents-spec.md
• decision-matrix-rerun.md
• update-back-rule.md
• living-instruction-file-spec.md
• stone-vs-opinion.md
• two-views-model.md