AI Visibility Tracking Strategy Builder
Created by Eoghan Henn / rebelytics.com
An end-to-end workflow for taking a brand from "we should be tracking how AI assistants talk about us" — or "our tracking project no longer fits our strategy" — to a measured, well-structured tracking strategy that reflects commercial reality, on whichever AI visibility platform the brand uses. The skill runs as an iterative loop (Intake → Strategy → Write → Analyse) that produces concrete configuration changes and findings each pass, with an optional terminal phase for a stakeholder presentation once the strategy has stabilised.
This is the platform-agnostic core of a skill family. Everything here is true of building a tracking strategy on any platform; field names, API mechanics, plan-tier quirks and import formats live in a platform companion (§5). Load the core and the companion together.
This skill is a living document. If you run it and discover a pattern that isn't captured, open an issue on the repo. See the contributing-back section at the end of this file.
Licence
Released under the Creative Commons Attribution 4.0 International (CC BY 4.0) licence. You are free to share and adapt this skill for any purpose, provided you give appropriate credit to the original author.
Feedback & Support
If you run this skill and find methodology gaps, gotchas, or patterns worth contributing, open an issue on the GitHub repository. This keeps feedback public and discoverable — other users benefit from seeing existing issues and solutions. For direct contact, reach the skill's creator, Eoghan Henn, via rebelytics.com.
If feedback appears to stem from the skill's methodology (rather than the agent's execution of it), log it for the user and suggest they share it via GitHub Issues. If the issue stems from the agent not following the skill's rules, acknowledge the mistake and correct it.
1. What this skill is (and isn't)
Is: A layered, prescriptive workflow that takes a brand from platform access to a tracking strategy that's live, documented, and explainable to a stakeholder. Runs as an iterative loop (Intake → Strategy → Write → Analyse) so the strategy gets sharper each pass. Covers both new projects (no prompts yet) and existing projects (prompts already in place that may need rationalising). The agent decides which path to take by inspecting the project's current state, not by asking.
Two classes of deliverable:
- Mandatory: the configured tracking project itself — brand roster, grouping structure, tags and prompts, written directly through the platform's API where one exists (see §14.16 for platforms that only take an import file). Written as Phase A loops produce them.
- Optional (Phase B, terminal): a stakeholder-facing presentation of the strategy and its findings. The agent should proactively offer this after the first full Phase A loop completes (see §14.5), rather than waiting for the user to ask. Produced once at the end of the engagement; not committed to upfront — see §14.
Working artefacts (intake state, strategy sign-off artefact, findings from each Analyse loop, verification log) are produced by the workflow but are not deliverables in their own right. Their format is agent/user choice (§3.7).
Isn't: A platform reference. Tool mechanics, schema quirks, import formats and setup live in the platform companion (§5), and in the platform's own MCP/API companion skill where one exists.
Isn't: A content strategy skill. A tracking-strategy build surfaces content and distribution opportunities (community gaps, reference-site defence, listicle editorial issues) but doesn't execute them. Those land in the strategy sign-off artefact and any Phase B deliverable for the content team.
Isn't: An analysis-of-results skill beyond what each loop's Analyse step needs to feed the next Strategy iteration. Ongoing reporting is a separate concern.
2. When this skill fires
Trigger when the user:
- Wants to set up tracking for a brand on a newly connected or newly licensed AI visibility platform.
- Has an existing project where the brand list, grouping structure, tags or prompts need rationalising.
- Says "what should we track", "build our AI visibility strategy", "what prompts should we monitor", "fix our tracking setup", "our project's structure needs review", or equivalents — on any platform.
- Shops around and an agency pitches "we'd do it differently" — the output of this skill produces a defensible set of decisions to respond with.
- Asks for a tracking strategy review ahead of a renewal or budget conversation.
- Wants a stakeholder presentation for an already-stable strategy (Phase B only — the Phase A loops have already produced the strategy in an earlier engagement).
- Is about to reuse one market's prompt allocation in another market (§9.2 governs; see the cross-market divergence check).
The skill works with any project state: zero prompts, 50 prompts still finding shape, or several hundred prompts that have grown organically. The workflow routes automatically based on what it finds.
3. Cross-cutting principles
These principles apply across every phase and loop of this skill. Re-read them at the start of each loop before planning work. They're the rules of the game — phase-specific rules (§§4, 8–15) defer to these when they conflict.
3.1 Instrument separation (branded vs non-branded)
Branded prompts and non-branded prompts measure different things. They're not two columns of one KPI — they're two instruments with different purposes. Never aggregate them into a single visibility score, never claim "the brand leads on branded prompts" as a finding (tautological), and always report them separately in analysis and strategy. If a combined metric is requested, produce it only with explicit disclosure that it's an anti-pattern.
What each instrument actually measures (name this in stakeholder outputs whenever the split appears):
- Non-branded prompts = commercial visibility scorecard. These measure whether the AI surface names the brand when asked commercial questions that don't contain the brand's name. This is the genuine discoverability signal and the only valid basis for competitive comparison.
- Branded prompts = reputation / representation instrument. These measure how the AI describes the brand when asked about it by name. Visibility on branded prompts is near-tautological (the brand name is in the prompt, so the brand nearly always appears in the response); sentiment and source authority on branded prompts carry real information.
Three hard rules that follow from the split:
- Headline metrics split side-by-side. Never report a blended visibility / share-of-voice figure as the primary number. Blended figures are dominated by branded prompts — the brand looks stronger than it is, and the non-branded competitive picture disappears.
- Competitive comparisons exclude branded prompts. Share of voice and position rankings only make sense on non-branded; branded prompts systematically favour whichever brand's name is in the prompt.
- Sentiment belongs to branded. Non-branded sentiment is mostly noise — most non-branded responses don't name the brand at all. Sentiment as a headline metric comes from the branded cohort.
For an intermediary (retailer, marketplace, distributor, agency) the split is three cohorts, not two — §4.10 and §9.7 carry the third case.
Few platforms have a native branded/non-branded concept. The standard implementation is a brand-mention tag set at prompt creation (§9.7); the companion says what the platform offers.
Phase B additional rule — the branded visibility number does not appear. In analysis contexts (§13), the split-side-by-side rule above is sufficient because the audience is reading the methodology carefully. In stakeholder contexts (§14), attention filters captions and leaves only the numbers behind — a branded visibility figure in the nineties next to a non-branded figure in the teens, even with a caption reading "branded measures reputation, not discovery", anchors the stakeholder on the big number as a success signal. Split-screen framing is not a sufficient guard against anchoring in a Phase B deliverable.
The Phase B rule is stricter than the Phase A rule:
- The branded visibility figure must not appear in any Phase B slide — not as a headline, not as a comparison, not in a caption, not as a sanity check, not on a methodology slide.
- The non-branded figure appears on its own, named clearly as the discoverability score (or equivalent stakeholder-register label, see §3.5).
- The branded cohort is acknowledged as "measured separately, reported on sentiment and source authority in the branded appendix" (or an equivalent one-liner); it is not presented with a percentage.
An unfiltered brand report silently inflates the own brand against competitors. A brand report with no cohort filter returns visibility computed across all prompts in scope — branded and non-branded combined. The blended figure is mathematically valid for any single brand individually, but as a competitive comparison it is structurally inflated for the own brand: on branded prompts the own brand scores near 100% by construction while every tracked competitor is measured on the same prompts and scores near 0% (the prompt isn't about them). The blended roster comparison is asymmetric by construction.
Hard rule — competitive comparisons. Any roster comparison (Phase B or Phase A) filters to non-branded only. Never report an unfiltered blend as a competitive headline.
Worked example (composition shape, not just numbers). An initial deck reported the own brand at "#1 in category visibility" on an unfiltered blend across roughly 70 prompts (a small branded slice plus a much larger non-branded slice). Recomputed on non-branded only, the own brand dropped several points and was essentially tied with the next-ranked rival. A different headline. The blend was the own brand's branded ~95% pulling its figure up, while every competitor's blend was held flat by their ~0% on the same branded prompts. Competitors had no equivalent inflation pathway.
This is the same family of error as the §14.2 "summing rates across brands" hard rule — applied to the instrument axis instead of the brand axis. Both rules generalise: aggregating across two distinct measurement instruments doesn't just produce uninterpretable numbers; when one instrument has structural asymmetries between the own brand and competitors (as branded prompts do), the aggregate silently inflates the own brand.
3.2 Provenance
Every claim in findings must be traceable to a platform read (which report, which filter, which date window, which cohort). If a claim can't be traced, it doesn't appear in findings. This is the discipline that keeps Analyse honest.
3.3 Caveats are symptoms, not cover
If a finding requires a caveat to be defensible ("this is strong, but note that the cohort was only 3 prompts"), the finding is weak — the caveat is a symptom that the evidence is thin. Either strengthen the evidence, reframe the finding, or drop it. Caveats shouldn't be used as a rhetorical escape hatch.
3.4 Calendar time is a resource
Measurement windows, cohort age, and signal maturity are real constraints, not formalities. "We wrote 25 deletes yesterday, let's analyse tomorrow" is wrong because the cohort hasn't stabilised. Plan loops around the signal, not the work rhythm. Communicate pacing expectations to the user at intake (§8).
Corollary — don't bake in cadences the skill hasn't itself validated. Specific review rhythms (30-day, 90-day, monthly) should be described as user-configurable defaults with the rationale exposed, not hard-coded as skill prescriptions. If the skill's validation timeframe hasn't covered multiple cycles at the prescribed cadence, treat it as a hypothesis the user can adjust, not a finding. Phrase cadence language accordingly ("Users typically review monthly; confirm what works for your team") rather than as a rule ("Reviews happen monthly").
3.5 Audience separation
The agent's internal reasoning, the user-facing sign-off, and the Phase B stakeholder deliverable are three different audiences with three different registers. Don't leak internal methodology language into stakeholder decks; don't leak stakeholder simplification into agent reasoning.
3.5.1 Structured-question register rules
Tool note. "AskUserQuestion" is the structured multi-choice user prompt used by Claude. Other agents expose equivalent primitives under different names. The rules below apply to any structured multi-choice user prompt, whatever the agent calls it.
The audience-separation principle (§3.5) applies to every user-facing surface, including AskUserQuestion payloads. Skill-internal labels exist for agent orientation; they must never appear as user-facing option labels or question text.
- Option labels must be outcome-focused, not process-focused. Prefer "I design the strategy, get your OK, then create everything in the tool" over "Phase A loop 1 — strategy + writes + verify".
- Never use skill-internal jargon in option labels or questions. Forbidden terms as user-facing labels: "Phase A / Phase B", "Ring 1/2/3 intake", "strategy sign-off artefact", "disposition framework", "prompt disposition", "loop 1", "rationalise", "Branch A/B", "quality gate".
- Option descriptions should describe what the user will experience, not what the skill's phases do. Good: "You end with a live, configured project collecting data from tomorrow." Bad: "Agent executes Write sub-phase after sign-off."
- When in doubt, assume the user has never seen the skill. This mirrors §3.5 for stakeholder decks but applies during the work, not after. The skill runs with the user inside it — they don't benefit from the skill-internal vocabulary the agent uses to navigate the workflow.
3.6 Label travel
Tags, container names, and strategy labels travel across deliverables and
sessions. A tag called gap-to-close in the tool needs to mean the same
thing in the strategy sign-off, in findings, and in a Phase B deck. Rename
with care, rename everywhere at once, and document what each label means
in the persisted intake state.
3.7 What, not how
This skill specifies what must happen, not how it's presented. Anything format-dependent (file names, sign-off medium, findings structure, state persistence mechanism) is agent/user discretion. The skill prescribes outcomes and gates; users choose their own formats.
3.8–3.10 Intake discipline
Intake shortcuts are the strongest failure mode this skill has seen; the
three named skip shapes, the recovery rule (a user-ask, never a
substitution) and the meaning of the intake "rings" are §3.8–3.10 and
live with the intake playbook — read references/phase-a-intake.md in
full at the start of every Intake step.
4. Core principles
4.1 Quality over quantity
Fewer well-chosen prompts always beat many marginal ones. Every prompt that doesn't earn its slot costs two things: platform credits or budget, and analytical noise that obscures signal from the prompts that matter.
The test each prompt must pass: "Does this prompt measure a question a real customer would ask on the way to a purchasing decision?" If the answer is "only loosely" or "for completeness", cut it. Resist the instinct to fill whatever ceiling the plan or the user names; propose the defensible number and document an optional tranche (§9.1).
4.2 Every prompt earns its slot
Slots are allocated based on data, not intuition. Demand signals (landing-page revenue, search-console click share, CPC × volume, AI-fanout patterns, observed AI citations, customer-voice logs) determine how many prompts a category deserves. Categories without external demand signals don't get slots just because the brand offers the product — they may be a content problem, not an AI visibility problem. Demand decides depth, though, not presence: where the budget is being sized rather than spent, every product the brand sells keeps a two-prompt floor, so a coverage gap cannot masquerade as a visibility finding (§9.1.0).
The converse also holds: every prompt is a claim that the brand could plausibly be the answer, so a prompt naming a service the brand does not offer is a structural zero, not a measurement — and that is checked against the brand's own catalogue for every prompt in the finished set, inherited ones included (§9.9, §15.4 gate 13). And the shares the strategy sets are filled as quotas and re-measured after every cut, because ranking a pool on value deletes the lowest-scoring instrument without leaving a trace (§9.1.4, §15.4 gate 12).
4.3 Automated sources first, user asks last
The agent must check what it can already access before asking the user for anything. The platform's own API or MCP, connected data MCPs (analytics, search console, SEO tools, commerce platforms, crawl tools), other loaded skills (brand dossiers, business context), and earlier conversation context all take precedence. The user is asked for manual input only as a batched fallback when no automated path exists. When asking, ask once, ask for the maximum useful batch, and never trickle questions across the session.
4.4 Data persistence
Any data the user provides manually — domain lists, market priorities, taxonomies, customer-voice samples, regulatory notes — is saved to the persistence store so the user never has to provide it twice. On subsequent runs, the agent reads the saved data first, surfaces it to the user, and asks only whether it needs refreshing. See §7 for the persistence mechanism (and note that the mechanism itself is agent/user choice per §3.7).
4.5 Prescriptive strategy, explicit overrides
The Strategy phase outputs a concrete recommendation the user accepts or rejects — not a menu of options with trade-offs. Each recommendation is paired with an explicit "Override this if…" block that lists the common departures. The user reads the recommendation, accepts, or calls out an override. No "which would you like" questions during the strategy phase.
4.6 Brand list reflects reality
Platform-suggested competitor lists are usually wrong — they skew toward information sites (forums, wikis, magazines) rather than actual commercial rivals. The real competitors are domains that actually get cited when AI models answer prompts in the brand's space. On day 0, derive the competitor list from the brand's own commercial knowledge plus any available SERP-competitor or citation data. On existing projects, derive it from the platform's domain and URL citation reports.
Three categories, not two. The brand roster has three shapes, not the binary "competitor vs marketplace noise":
- Commercial competitor — a distinct business the own brand is fighting for the same customer's wallet. Add to the roster as a competitor; it will drive share-of-voice, gap reports, and competitive narrative.
- Assortment brand — a brand stocked by the own retailer (it appears as a product line or brand page on the own site). Retailing the brand does not make it a competitor; conflating the two inflates competitor counts and corrupts gap analysis. Assortment brands can still matter for visibility tracking (a user searching for the brand may land on the retailer) but should be tagged as assortment so the Strategy-phase output surfaces them distinctly.
- Marketplace / generic noise — large marketplaces, shopping engines, generic directory pages. Not a competitor; not stocked; just co-retrieval noise.
§9.3 codifies the classification step to run before adding any brand to the roster, and how assortment brands map to the grouping structure.
4.7 Tags serve analysis, not categorisation
A tag taxonomy should be 2–3 dimensional (intent × funnel × category is
a common shape) and total 15–25 tags. More than that and consistency
breaks down. Every prompt should carry at least one tag from each
dimension. Avoid tag-dimension duplication (a transactional tag and a
funnel:decision tag are the same signal — pick one).
4.8 Containers mirror business structure — cardinality decides the rest
Every platform offers some single-valued grouping field (a brand, a topic, a cluster — the companion names it) and usually a multi-valued tag field. The single-valued container is the coarse grouping used for dashboards and reporting. It should map to the brand's internal structure so stakeholders can find their area. Don't duplicate categorisation between containers and tags — let containers carry business structure and tags carry cross-cutting dimensions.
Cardinality is what settles the container-versus-tag question. A single-valued field can only ever hold the one dimension that genuinely partitions the set — so the container must carry the dimension the budget is allocated against and that stakeholders own internally, normally the commercial category. Every dimension that overlaps categories (funnel stage, intent, seasonality, regulatory sensitivity, diagnostic status, brand-mention type, assortment brand) has to live in the multi-valued field or be lost, because multi-membership is expressible nowhere else. The question is never "containers or tags" — it is which dimension goes in which field, and the answer follows from how many values the field holds. Where a nested sub-level exists, it must subdivide its parent, never act as a second independent axis: two orthogonal axes nested in one field cannot be filtered apart afterwards. §9.4 works this through for the three field shapes platforms actually expose.
5–8 containers covers most single-market e-commerce projects; more than 10 usually signals either multiple markets mashed together or containers acting as tags. Containers are categories, not brand names — a competitor's name belongs on the brand roster, never as a container.
4.9 Engine coverage matches the audience
Every tracked engine multiplies chat count and cost. Match engines to where the audience actually asks — and don't default to "all engines". Some platforms gate the engine set by plan tier; there the question flips from "which engines" to "am I getting maximum signal from the engines I do have?" — a prompt-quality and brand-detection-hygiene question. The companion says how to detect the plan's engine set (§9.6).
4.10 Branded, other-brand and non-branded prompts are different KPIs
This section applies the instrument-separation principle (§3.1) to KPI reporting specifically — see §3.1 for the underlying rule.
Rolling branded-prompt visibility into overall visibility inflates the headline number. A project with 10 branded prompts that score near 100% and 90 non-branded prompts at 10% looks like it has ~20% visibility when the real "unprompted discovery" visibility is 10%. Branded and non-branded metrics must be reported separately. Keep branded prompts small in count (5–10) and consistently tagged so filters work.
For an intermediary the split is three cohorts, not two. Before adopting any brand-mention taxonomy, ask what fraction of the tracked entity's value comes from entities it does not own. For a single-brand business the answer is "none" and the binary holds. For a retailer, marketplace, distributor or agency the answer is "most of it" — an intermediary's whole business is other people's brands — and the case the binary loses is the main case:
- Own brand named — near-guaranteed mention; the reputation instrument.
- Another brand named, own brand not — a rival, or an assortment brand the retailer stocks (§4.6). The own brand is not guaranteed a mention here; it has to be surfaced as a place to buy that brand. A hard signal, and for an intermediary frequently the most actionable one.
- No brand named — the pure discoverability instrument.
Tag on who is named in the prompt text. Whether the named brand competes or supplies does not change the tag: the same company is often both, and only "who is named" has a stable answer. Leaving the middle case undefined guarantees inconsistent labelling — handed a two-value scheme, parallel authors tag the identical construction both ways, each defensibly. §9.7 carries the decision and the reporting arithmetic lives in §13.
4.11 Sign-off is the audit trail; verification is the closure
Where the platform has a write API, the skill writes directly to the project rather than producing an intermediate operations list. The strategy sign-off artefact — produced before writes begin — is the audit trail. User sign-off happens before writes begin; verification happens after, by re-reading the platform state and reconciling it against the plan (§12). Where the platform only takes an import file, the validated file is the write and the post-import re-read is the verification (§14.16).
4.12 Fresh prompts need time
Prompts start collecting data within roughly a day on most platforms. The default iteration pace is daily — the next Analyse loop can run ~24 hours after writes. For trend-level analyses (share-of-voice shifts, sentiment drift), 7–14 days produces more stable signal. Don't run Analyse in under 24 hours — the data won't be there yet. See §12.7 for the full measurement-window guidance and §3.4 on why calendar time is a first-class resource.
4.13 Wildcard over granular
Each prompt should represent a cluster of related queries, not a single exact phrase. AI conversations are personalised — the same person might ask "best trail running shoes for beginners", "which running shoes should a beginner buy for trails" or "I'm starting trail running, what shoes do I need?" These all map to the same intent. Track the representative wildcard prompt; don't burn prompt slots on variations of the same question.
The non-determinism counterargument, and why it doesn't break this. Public advice sometimes recommends running dozens of phrasing variations per priority query because individual AI responses are non-deterministic. That addresses a different layer. Prompt-set composition (what the wildcard principle governs) is how a finite budget is spread across the intents the business cares about; measurement reliability (what the non-determinism argument addresses) is how confidently any single prompt's score reflects reality, given stochastic outputs. Mainstream platforms handle non-determinism temporally — they re-run each prompt on a schedule and aggregate across runs, so a prompt's score is already a distribution over many samples. Enumerating phrasing variations at the prompt-set level is largely redundant with what the tool already does, and it buys depth at the cost of breadth. The one genuine exception is surgical precision on a very small number of strategically critical queries — a supplementary exercise, never the default framing. If a stakeholder pushes back with the non-determinism argument, the reply is: the concern is real, it's handled at the measurement layer, and the prompt set is optimised for breadth because that's where the business value is.
4.14 Commercial intent filter
Every prompt should be close to a purchasing decision. The test: "If someone asks an AI model this question and gets an answer that includes our brand, would that plausibly lead to a sale or enquiry?" Include "best [product] for [need]", "how do I choose [product]", "[brand] vs [competitor]", "what does [service] cost". Exclude "what is [concept]", "history of [topic]", "[acronym] meaning" — unless the business model is inherently educational, in which case training-intent queries qualify. The deliberate exception is a small, tagged band of owned-territory probes on informational ground the brand is already cited for (§9.9). A second, equally deliberate exception is a small trust / reviews block of branded prompts per market (§9.9), included for source diagnosis rather than measurement — it reveals which third-party review pages and profiles the engines pull for the brand — and excluded from or reported separately in the headline visibility figure like any other branded prompt.
4.15 Conversational framing
Prompts should sound like how a person talks to an AI assistant, not like a search-engine keyword. Good: "How do I get my bike insured for racing?" Bad: "bike insurance racing".
4.16 Market-specific adaptation
The same business may need different prompts per market. Build each language/market as an independent prompt set: share the container structure, let the individual prompts diverge on local demand, and never inherit an allocation from a sibling market on structural grounds — run the cross-market divergence check in §9.2 first.
5. Relationship to other skills
| Skill | Role |
|---|---|
| Platform companion | Required alongside this skill whenever the platform has one. Holds the platform's field names, API/MCP mechanics, plan-tier quirks, import format and platform-only patterns, as "§N — implementation" parts keyed to this skill's section numbers. Published companions: peec-ai-tracking-strategy-builder (Peec AI), sistrix-tracking-strategy-builder (SISTRIX AI visibility / custom prompt tracking), searchviu-tracking-strategy-builder (searchVIU AI visibility) and llmpulse-tracking-strategy-builder (LLM Pulse). For any other platform, run this skill alone and treat its import or API contract as the companion's job. |
| Platform MCP/API companion (varies) | Where the platform's tool surface has its own skill (e.g. peec-ai-mcp, sistrix-mcp, searchviu-mcp), the platform companion names it. Where the platform's MCP server carries its own tool descriptions and no separate skill exists (LLM Pulse), the companion points at the server instead. This skill consumes its reads; it doesn't restate tool mechanics. |
| External-data skills (varies) | Any skill the user has for pulling search-console data, SEO-tool data, crawl data, analytics data or AI-citation data. This skill consumes their output; doesn't reinvent them. |
| Brand-context skills (varies) | Any skill the user has loaded that holds accumulated knowledge about the brand being tracked (brand dossier, project-context skill, or similar — however it's named in the agent's environment). Contains prior-captured context (markets, TLDs, revenue profile, regulatory notes) that bypasses intake questions. Always check. |
How the split works in practice. Section numbers are global across the family. When a section here says "the companion gives the platform's mechanics", the companion holds a part with the same number ("§9.3 — Peec implementation"). Read the core section first, then the companion part. Companions never restate a core rule; if a companion appears to contradict this skill, this skill wins for method and the companion wins for facts about the platform. The same holds for enumerated lists: a companion points at the §15.4 validation gates by the core's gate number and never copies the list, so that a gate added here shows up in the companion as a missing number rather than a silent omission.
6. Workflow overview
The workflow runs as Phase A (iterative) plus an optional Phase B (terminal). Phase A is where all the configuration work happens; Phase B is an opt-in stakeholder deliverable at the end.
Phase A — iterative loop
Each loop runs four sub-phases in sequence:
- Intake (§8). Gather everything the strategy needs. Three concentric rings on the first loop, run as parallel paths (§3.10): automated tools (Ring 1), a baseline path as the floor (Ring 2), and a batched user-provided data ask (Ring 3). Outputs a populated intake state, persisted per §7.
- Strategy (§9). Convert intake into a prescriptive recommendation with explicit override callouts. User signs off (format agent/user choice). Outputs the strategy sign-off artefact.
- Write (§12). Execute the signed-off strategy as direct writes to the project via the platform's API — or as a validated import file where that is the platform's contract. Outputs a configured project and a verification log.
- Analyse (§13). Read what the platform's data actually says after the writes have settled. Produce findings (format agent/user choice) that feed the next Strategy iteration.
Loop-awareness. The first loop runs the full three-ring intake (§8). Subsequent loops typically run a platform-only intake refresh — re-reading current project state — unless findings or the user surface a gap that needs external data, in which case Ring 2 or Ring 3 is re-entered for that specific gap. Strategy, Write, and Analyse run every loop, with scope narrowed to what changed.
Termination. Phase A ends when findings no longer produce material action for the next Strategy iteration — the strategy has stabilised. That's also when Phase B becomes relevant.
Phase B — optional, terminal
Once Phase A has stabilised, the agent proactively offers a stakeholder-facing presentation of the strategy and its findings (see §14.5 for the timing rules). Phase B is a single pass from existing Phase A artefacts — it doesn't re-do any Phase A work. See §14 for the full rules, including why Phase B is intentionally terminal rather than mid-engagement.
Routing by project state
Phase A's Write sub-phase routes on project state:
- New project (zero prompts at first loop) → first Write is pure creates.
- Existing project → first Write uses the prompt disposition framework (§10) — six buckets covering what to keep, retag, reframe, or remove.
Subsequent loops always run in existing-project mode; the disposition framework is reapplied to the prompts that have been in place long enough to have meaningful data.
Pacing disclosure at intake
At the start of the first Intake loop, the agent surfaces the iterative model and the iteration cadence to the user:
- The default pace is daily iterations — most platforms process new prompts within about 24 hours (the companion states the platform's own figure), so the next Analyse loop can run the following day (§12.7).
- For trend-level analyses (share-of-voice shifts, sentiment drift), a 7–14 day window produces more stable signal and can be used when the question requires it.
- The number of loops is driven by findings, not by a preset timeline.
- Once the strategy has stabilised across a few iterations, the agent can produce a polished stakeholder-facing presentation of the strategy and its findings (Phase B, §14). That comes later — first the tracking needs to be right. Phase B is not committed to at intake, but the user should know the capability exists.
This framing sets expectations before the user has anchored on a linear "do it once and we're done" model.
6a. Section map — where to read what
This skill uses progressive disclosure: this file holds the mental model
and the rules that apply on every invocation; the phase playbooks live in
references/ and are loaded when their phase runs. Section numbers are
global across the family — a cross-reference like "§13.6" resolves via
this map, and the platform companion's map says where its
"§N — implementation" parts live. Every load trigger below is mandatory,
not optional: running a phase without reading its file first is the same
failure shape as skipping an intake ring (§3.8).
| File | Sections | Load trigger |
|---|---|---|
references/data-persistence.md |
§7 | When initialising or resuming the project workspace, and before writing any loop artefact |
references/phase-a-intake.md |
§3.8–3.10, §8 (incl. §8.5 data-source handling) | In full, at the start of every Intake step, before asking the user for anything |
references/phase-a-strategy.md |
§9–§10 (incl. §9.1.0 allocation direction, §9.1.4 fill-by-quota and the answer-shape classifier, §9.9 prompt authoring, product detectors, the disambiguation pass and the trust / reviews block) | Before drafting or revising any strategy recommendation, before authoring or delegating prompts, before finalising any prompt set, and before sign-off |
references/allocate.py |
§9.1.2 | Run (do not re-derive) whenever the damped revenue-share method leads the allocation |
references/pattern-library.md |
§11 | Skim its contents list during every Strategy and every Analyse step; read any pattern whose symptom matches |
references/write-and-analyse-principles.md |
§12–§13 (principles, incl. §12.7 measurement windows and run frequency, §13.5 per-answer share) | Before the first Write of a loop and at the start of every Analyse step, and before quoting any visibility figure or making any run-frequency decision — then load the companion's §12/§13 for the platform mechanics |
references/phase-b-deliverable.md |
§14 (incl. §14.16 configuration deliverable) | When Phase B is triggered or offered — plus the §14.5 timing check-in at every loop close; §14.16 whenever the platform takes an import file rather than API writes |
references/quality-gates.md |
§15 (incl. §15.4 merged-set validation gates) | At every phase transition; the relevant gate must pass before Write, Analyse, or Phase B begins, and §15.4 before any prompt set is written or delivered |
Then the companion's section map for the platform parts of §7–§15.
16. Output deliverables
16.1 Mandatory deliverable
The configured tracking project itself. Brand roster, containers, tags and prompts are live on the platform as the result of the Phase A Write sub-phase — written through the API, or delivered as the validated import files of §14.16 where that is the platform's contract. No intermediate operations list or execution log — the write is the execution, and verification (§12.5) is the closure.
Each Phase A loop adds to or refines the configured project. The "deliverable" grows with the loops; there is no single point at which it's "done" until the user decides Phase A has stabilised.
16.2 Optional deliverable
Phase B stakeholder presentation. Produced once, at the end of the engagement, if the user requests it (§14). Format is whatever suits the audience — a deck, a narrative document, a one-page summary, or something else. The that is optional; the what is covered by §14.
16.3 Working artefacts
Working artefacts are produced by the workflow but are not deliverables in the client-facing sense. Their format and storage mechanism are agent/user choice (§3.7). The skill requires they exist; it does not require a specific file layout.
- Intake state. Whatever format persists the brand config, markets, competitors, regulatory context, and data-source inventory (§7.2 is one example schema).
- Strategy sign-off artefact. The per-loop confirmation of accepted recommendations — a markdown document, a structured chat message, an interactive list, or a handoff doc (§9.8).
- Findings. The per-loop ou
…(truncated)