GTM Systems Audit
Score the four layers, find the constraint, and sequence the fixes. Run this before building anything.
When to use
- Day one of an engagement or a new role
- Somebody wants to buy or build something and nobody has diagnosed why
- Quarterly, as a standing review
- After a stack change, a migration, or a team change
Inputs
- Reads: anything already in
workspace/
- Needs from user: CRM access or a screenshot of the object counts, a list of tools and their annual cost, the current scoring model if one exists, sample of recent enrichment output, and one recent report everybody argues about
Ask for these in one batch. Do not start scoring with half of them.
Workflow
1. Score each layer 1 to 5
Score against the anchor, not against a feeling. A 3 is the honest default.
Data plane
| Score |
What it looks like |
| 1 |
No target list. Reps build their own. Duplicates everywhere |
| 2 |
A list exists, exported once, already stale. No scoring |
| 3 |
Maintained list, some enrichment, a score nobody trusts |
| 4 |
Fit and intent scored separately, thresholds fitted, refresh queue running |
| 5 |
The above plus cost per usable record tracked and a passing discrimination test |
Agent plane
| Score |
What it looks like |
| 1 |
Nothing built. AI use is individuals pasting into chat |
| 2 |
Someone built something. It works on their machine only |
| 3 |
Shipped, some adoption, guardrails are verbal |
| 4 |
Context portfolio exists, guardrails are enforced in the tool, adoption above 60% |
| 5 |
The above plus versioned distribution and tracked correction rate |
Quality plane
| Score |
What it looks like |
| 1 |
No quality bar written down. "We know it when we see it" |
| 2 |
A checklist exists somewhere. Applied inconsistently |
| 3 |
Rubrics exist, checked in the same context that produced the work |
| 4 |
Independent judge, hard gates, every run logged including failures |
| 5 |
The above plus calibration runs and drift detection |
Systems of record
| Score |
What it looks like |
| 1 |
Two systems both claim to be the source of truth |
| 2 |
One CRM, unenforced stage definitions, attribution by argument |
| 3 |
Definitions exist, mostly followed, one reporting layer |
| 4 |
Metrics dictionary with owners, routing SLAs measured, sync monitored |
| 5 |
The above plus stack spend inside the healthy band and utilization reviewed |
Use docs/metrics.md for the numeric checks behind each anchor.
2. Run the four fast diagnostics
These take under an hour combined and they are worth more than the interviews.
Duplicate check. Export accounts, normalize domains, count collisions. Over 8% and Layer 1 is capped at 2 no matter what else is true.
Score discrimination. Take last year's closed-won and closed-lost, score both with the current model, compare rates. Under 2x lift and the score is decoration.
Definition test. Ask three people to define one metric independently, in writing. Three different answers means Layer 4 is capped at 2.
Adoption check. For anything already built, count weekly active users against intended users. Under 30% means the build failed regardless of quality.
3. Name one constraint
Not a list of problems. One constraint, which is the lowest-scoring layer that everything else waits on.
Total interpretation:
| Total /20 |
Reading |
| 4 to 8 |
Foundational. Do not build agents. Fix the data model and definitions |
| 9 to 13 |
Targeted fixes. One layer is dragging the rest |
| 14 to 17 |
Ready to build. Instrument as you go |
| 18 to 20 |
Optimize and expand scope |
4. Sequence with gates, not dates
Each phase has an exit condition. No phase starts until the previous one passes.
PHASE 1 <objective>
Exit gate: <measurable condition>
PHASE 2 <objective>
Exit gate: <measurable condition>
PHASE 3 <objective>
Exit gate: <measurable condition>
Three phases maximum. Do not plan past the third, because the second one will change what the third should be.
5. Say what not to do
The most valuable section. List the things the team is about to do that this audit says no to, with the reason. An audit with no "do not do this" list gets treated as a suggestion.
Output
- Writes:
workspace/audit/systems-audit.md
- Uses:
templates/systems-audit-scorecard.md
- Prints: four scores with the evidence for each, the four diagnostic results, the named constraint, the three gated phases, and the do-not-do list
Rules & quality bar
- Run the four diagnostics. Interviews tell you what people believe. The diagnostics tell you what is true
- Score against anchors, never against impressions
- One constraint. A list of eight problems is not a diagnosis
- Gates, not dates. Calendar plans slip and get renegotiated. Gates do not
- Never recommend a purchase in this skill. Route to
martech-stack-audit if buying is genuinely the answer
- Include the do-not-do list. Always
Related skills
- Routes to: whichever layer scored lowest
- Hands off to:
data-model-and-sync in most cases, because most audits land there
- See also:
docs/architecture.md, docs/metrics.md
1---2name: gtm-systems-audit3description: Use at the start of any engagement, or when the user says "audit our systems", "where do we start", "what's broken in our GTM", "assess our stack", "we need to fix our operations", "new client onboarding". Scores all four layers 1 to 5, names the single constraint, and returns a sequenced fix plan with gates. Writes workspace/audit/systems-audit.md. Scores all four layers. Not a tool-by-tool inventory of what you pay for, which is `martech-stack-audit`.4---56# GTM Systems Audit78Score the four layers, find the constraint, and sequence the fixes. Run this before building anything.910## When to use11- Day one of an engagement or a new role12- Somebody wants to buy or build something and nobody has diagnosed why13- Quarterly, as a standing review14- After a stack change, a migration, or a team change1516## Inputs17- Reads: anything already in `workspace/`18- Needs from user: CRM access or a screenshot of the object counts, a list of tools and their annual cost, the current scoring model if one exists, sample of recent enrichment output, and one recent report everybody argues about1920Ask for these in one batch. Do not start scoring with half of them.2122## Workflow2324### 1. Score each layer 1 to 52526Score against the anchor, not against a feeling. A 3 is the honest default.2728**Data plane**2930| Score | What it looks like |31|---|---|32| 1 | No target list. Reps build their own. Duplicates everywhere |33| 2 | A list exists, exported once, already stale. No scoring |34| 3 | Maintained list, some enrichment, a score nobody trusts |35| 4 | Fit and intent scored separately, thresholds fitted, refresh queue running |36| 5 | The above plus cost per usable record tracked and a passing discrimination test |3738**Agent plane**3940| Score | What it looks like |41|---|---|42| 1 | Nothing built. AI use is individuals pasting into chat |43| 2 | Someone built something. It works on their machine only |44| 3 | Shipped, some adoption, guardrails are verbal |45| 4 | Context portfolio exists, guardrails are enforced in the tool, adoption above 60% |46| 5 | The above plus versioned distribution and tracked correction rate |4748**Quality plane**4950| Score | What it looks like |51|---|---|52| 1 | No quality bar written down. "We know it when we see it" |53| 2 | A checklist exists somewhere. Applied inconsistently |54| 3 | Rubrics exist, checked in the same context that produced the work |55| 4 | Independent judge, hard gates, every run logged including failures |56| 5 | The above plus calibration runs and drift detection |5758**Systems of record**5960| Score | What it looks like |61|---|---|62| 1 | Two systems both claim to be the source of truth |63| 2 | One CRM, unenforced stage definitions, attribution by argument |64| 3 | Definitions exist, mostly followed, one reporting layer |65| 4 | Metrics dictionary with owners, routing SLAs measured, sync monitored |66| 5 | The above plus stack spend inside the healthy band and utilization reviewed |6768Use `docs/metrics.md` for the numeric checks behind each anchor.6970### 2. Run the four fast diagnostics7172These take under an hour combined and they are worth more than the interviews.7374**Duplicate check.** Export accounts, normalize domains, count collisions. Over 8% and Layer 1 is capped at 2 no matter what else is true.7576**Score discrimination.** Take last year's closed-won and closed-lost, score both with the current model, compare rates. Under 2x lift and the score is decoration.7778**Definition test.** Ask three people to define one metric independently, in writing. Three different answers means Layer 4 is capped at 2.7980**Adoption check.** For anything already built, count weekly active users against intended users. Under 30% means the build failed regardless of quality.8182### 3. Name one constraint8384Not a list of problems. One constraint, which is the lowest-scoring layer that everything else waits on.8586**Total interpretation:**8788| Total /20 | Reading |89|---|---|90| 4 to 8 | Foundational. Do not build agents. Fix the data model and definitions |91| 9 to 13 | Targeted fixes. One layer is dragging the rest |92| 14 to 17 | Ready to build. Instrument as you go |93| 18 to 20 | Optimize and expand scope |9495### 4. Sequence with gates, not dates9697Each phase has an exit condition. No phase starts until the previous one passes.9899```100PHASE 1 <objective>101 Exit gate: <measurable condition>102PHASE 2 <objective>103 Exit gate: <measurable condition>104PHASE 3 <objective>105 Exit gate: <measurable condition>106```107108Three phases maximum. Do not plan past the third, because the second one will change what the third should be.109110### 5. Say what not to do111112The most valuable section. List the things the team is about to do that this audit says no to, with the reason. An audit with no "do not do this" list gets treated as a suggestion.113114## Output115- Writes: `workspace/audit/systems-audit.md`116- Uses: `templates/systems-audit-scorecard.md`117- Prints: four scores with the evidence for each, the four diagnostic results, the named constraint, the three gated phases, and the do-not-do list118119## Rules & quality bar120- **Run the four diagnostics.** Interviews tell you what people believe. The diagnostics tell you what is true121- **Score against anchors,** never against impressions122- **One constraint.** A list of eight problems is not a diagnosis123- **Gates, not dates.** Calendar plans slip and get renegotiated. Gates do not124- **Never recommend a purchase in this skill.** Route to `martech-stack-audit` if buying is genuinely the answer125- **Include the do-not-do list.** Always126127## Related skills128- Routes to: whichever layer scored lowest129- Hands off to: `data-model-and-sync` in most cases, because most audits land there130- See also: `docs/architecture.md`, `docs/metrics.md`
guerrilla2799