Founder Review
Inputs
$request: Optional plan name, direct plan path, or mode override such as auth-system, .ulpi/plans/foo.md, --quick, or foo --full.
Goal
Catch plan defects before agents execute them: phantom file paths, markdown/JSON drift, hidden prerequisites, bad task boundaries, unsafe rewrites, missing tests, and overconfident execution graphs.
Step 0: Resolve the target plan and review mode
Locate the plan to review:
- if
$request names a plan, resolve .ulpi/plans/<name>.md
- if
$request provides a direct path, use it
- if no target is provided, select the most recently modified
.md plan under .ulpi/plans/
Then:
- read the full markdown plan
- read the companion JSON file if present
- extract title, mode, task count, task IDs, file paths, and dependency summary
- determine whether review mode is
FULL or QUICK
Load references/review-rubric.md for:
- mode selection rules
- section requirements per mode
- verdict criteria
Stop early if:
- no plan exists
- the markdown plan cannot be read
- the review target is too small for formal founder review and should just be executed directly
Success criteria: The target plan and review mode are explicit before detailed review starts.
Step 1: Run the codebase reality check
Verify the plan against the actual repository:
- every
filesToModify path exists
- every
filesToCreate path has a valid parent path or prior creator task
- markdown and JSON agree on task IDs, file paths, and execution meaning
- claimed reuse actually exists locally
- the plan is not building functionality that already exists
- if the plan introduces a public SQL, API, or CLI surface, the plan pins the real signature or example rather than hand-waving it
Load references/review-checklist.md and complete Section 1 in full.
Success criteria: The plan's file paths, reuse claims, and artifact consistency are reality-checked against the repo.
Step 2: Challenge scope and strategy
Review:
- whether plan mode matches actual scope
- whether every task serves the stated goal
- whether non-goals are credible
- whether the ship cut is believable
- whether hidden prerequisites or scope creep exist
Load references/review-checklist.md and complete Section 2.
Success criteria: Scope, mode, and minimum shippable cut are defensible.
Step 3: Audit architecture, contracts, and dependencies
Review:
- architecture diagram completeness
- task-to-diagram consistency
- producer/consumer contract clarity
- dependency JSON realism
- execution-summary consistency
- ownership of shared wiring points such as routers, registries, manifests, startup hooks, or export barrels
- semantic safety for rewrites, planners, optimizers, filters, or query composition
Load references/review-checklist.md and complete Section 3.
If review mode is QUICK, skip directly to Step 6 after this step.
Success criteria: The plan's architecture and dependency story is coherent and executable.
Step 4: Audit risk, recovery, and degraded-mode safety
For FULL review only, challenge:
- failure-mode coverage
- mitigation quality
- rollback or containment strategy
- degraded-mode semantics
- malformed-output or upstream-failure handling
Load references/review-checklist.md and complete Section 4.
Success criteria: Critical risks are covered by actionable mitigations rather than vague warnings.
Step 5: Audit test coverage and execution feasibility
For FULL review only, challenge:
- completeness of the test coverage map
- appropriateness of test types
- edge-case coverage in acceptance criteria
- security-sensitive path coverage
- DAG efficiency
- agent matching
- effort calibration
- P0 foundation honesty
- write-scope collisions or hidden overlap
Load references/review-checklist.md and complete Sections 5 and 6.
Success criteria: The plan is testable, executable, and not obviously over-constrained or under-specified.
Step 6: Score sections and deliver the verdict
Use references/review-rubric.md to:
- score each reviewed section as
PASS, WARN, or FAIL
- classify findings as
BLOCK, CONCERN, or OBSERVATION
- determine
APPROVE, REVISE, or REJECT
Then render a concise founder-review report that includes:
- plan title and mode
- verdict
- section summary table
- blocking issues
- concerns
- observations
- recommended plan changes when revision is required
Do not ask the user procedural questions at the end of the review. Deliver the audit and let the main thread decide what happens next.
Success criteria: The verdict follows directly from evidence-backed findings and the rubric.
Guardrails
- Do not modify the plan file as part of this skill.
- Do not rubber-stamp a non-trivial plan.
- Do not invent findings; a clean section can legitimately pass.
- Do not add
disable-model-invocation; this skill should remain callable when the user asks for plan review.
- Do not add
paths:; this is a generic review workflow.
- Do not keep giant examples, quality scorecards, or gate tables inline in
SKILL.md.
- Do not use
AskUserQuestion as a substitute for a verdict.
- Do not forget that this skill runs best as a forked review workflow with separate reasoning budget.
When To Load References
references/personality.md
Use at session start for review persona, traits, and communication style.
references/review-checklist.md
Use for section-by-section audit checks and evidence requirements.
references/review-rubric.md
Use for mode selection, verdict rules, section scoring, and finding classification (BLOCK/CONCERN/OBSERVATION gate tables).
Output Contract
Report:
- reviewed plan and selected mode
- verdict
- section status summary
- blocking issues
- concerns
- observations
- required plan changes when applicable
1---
2name: plan-founder-review
3version: 2.0.0
4description: |
5 Gate an implementation plan BEFORE agents execute it — catch phantom paths and false reuse, not typos:
6 read the full plan markdown and its companion JSON, reality-check every file path and reuse claim
7 against the actual repo, challenge scope, architecture, the dependency graph, and risk and test
8 coverage, then score each section and render an APPROVE / REVISE / REJECT verdict — runs as a forked
9 review workflow with its own reasoning budget, isolated from the main execution flow. The verdict
10 follows only from evidence-backed findings: never rubber-stamps a non-trivial plan and never invents
11 findings, but a genuinely clean section is allowed to pass; read-only, never edits the plan. Use when
12 the user asks to review or gate a plan before execution.
13allowed-tools:
14 - Read
15 - Glob
16 - Grep
17 - Bash
18context: fork
19agent: general-purpose
20effort: high
21argument-hint: "<plan-name|plan-path> [--quick|--full]"
22arguments:
23 - request
24when_to_use: |
25 Use when the user asks to review a plan, check a DAG plan, or gate an implementation plan before
26 execution. Examples: "review my plan", "/plan-founder-review auth-system", "/plan-founder-review
27 --quick", or "founder review this plan before we run agents". Do NOT use to generate the plan (that's
28 plan-to-task-list-with-dag), to execute its tasks, or to review code changes (that's find-bugs) —
29 this audits the plan artifact, nothing else.
30metadata:
31 source: https://skillmd.com/skills/ulpi-io/plan-founder-review
32---
33
34<EXTREMELY-IMPORTANT>
35This skill is a read-heavy plan-audit workflow. Non-negotiable rules:
361. Read the full plan markdown and the companion JSON if it exists.
372. Verify plan claims against the actual codebase instead of trusting the plan text.
383. Challenge missing dependencies, phantom paths, dishonest reuse claims, and semantic gaps.
394. Deliver a verdict only after the required review sections are complete.
405. Keep the invocation body focused on the review loop; load detailed checks and verdict rules from references.
41</EXTREMELY-IMPORTANT>
42
43# Founder Review
44
45## Inputs
46
47- `$request`: Optional plan name, direct plan path, or mode override such as `auth-system`, `.ulpi/plans/foo.md`, `--quick`, or `foo --full`.
48
49## Goal
50
51Catch plan defects before agents execute them: phantom file paths, markdown/JSON drift, hidden prerequisites, bad task boundaries, unsafe rewrites, missing tests, and overconfident execution graphs.
52
53## Step 0: Resolve the target plan and review mode
54
55Locate the plan to review:
56
57- if `$request` names a plan, resolve `.ulpi/plans/<name>.md`
58- if `$request` provides a direct path, use it
59- if no target is provided, select the most recently modified `.md` plan under `.ulpi/plans/`
60
61Then:
62
63- read the full markdown plan
64- read the companion JSON file if present
65- extract title, mode, task count, task IDs, file paths, and dependency summary
66- determine whether review mode is `FULL` or `QUICK`
67
68Load `references/review-rubric.md` for:
69
70- mode selection rules
71- section requirements per mode
72- verdict criteria
73
74Stop early if:
75
76- no plan exists
77- the markdown plan cannot be read
78- the review target is too small for formal founder review and should just be executed directly
79
80**Success criteria**: The target plan and review mode are explicit before detailed review starts.
81
82## Step 1: Run the codebase reality check
83
84Verify the plan against the actual repository:
85
86- every `filesToModify` path exists
87- every `filesToCreate` path has a valid parent path or prior creator task
88- markdown and JSON agree on task IDs, file paths, and execution meaning
89- claimed reuse actually exists locally
90- the plan is not building functionality that already exists
91- if the plan introduces a public SQL, API, or CLI surface, the plan pins the real signature or example rather than hand-waving it
92
93Load `references/review-checklist.md` and complete Section 1 in full.
94
95**Success criteria**: The plan's file paths, reuse claims, and artifact consistency are reality-checked against the repo.
96
97## Step 2: Challenge scope and strategy
98
99Review:
100
101- whether plan mode matches actual scope
102- whether every task serves the stated goal
103- whether non-goals are credible
104- whether the ship cut is believable
105- whether hidden prerequisites or scope creep exist
106
107Load `references/review-checklist.md` and complete Section 2.
108
109**Success criteria**: Scope, mode, and minimum shippable cut are defensible.
110
111## Step 3: Audit architecture, contracts, and dependencies
112
113Review:
114
115- architecture diagram completeness
116- task-to-diagram consistency
117- producer/consumer contract clarity
118- dependency JSON realism
119- execution-summary consistency
120- ownership of shared wiring points such as routers, registries, manifests, startup hooks, or export barrels
121- semantic safety for rewrites, planners, optimizers, filters, or query composition
122
123Load `references/review-checklist.md` and complete Section 3.
124
125If review mode is `QUICK`, skip directly to Step 6 after this step.
126
127**Success criteria**: The plan's architecture and dependency story is coherent and executable.
128
129## Step 4: Audit risk, recovery, and degraded-mode safety
130
131For `FULL` review only, challenge:
132
133- failure-mode coverage
134- mitigation quality
135- rollback or containment strategy
136- degraded-mode semantics
137- malformed-output or upstream-failure handling
138
139Load `references/review-checklist.md` and complete Section 4.
140
141**Success criteria**: Critical risks are covered by actionable mitigations rather than vague warnings.
142
143## Step 5: Audit test coverage and execution feasibility
144
145For `FULL` review only, challenge:
146
147- completeness of the test coverage map
148- appropriateness of test types
149- edge-case coverage in acceptance criteria
150- security-sensitive path coverage
151- DAG efficiency
152- agent matching
153- effort calibration
154- P0 foundation honesty
155- write-scope collisions or hidden overlap
156
157Load `references/review-checklist.md` and complete Sections 5 and 6.
158
159**Success criteria**: The plan is testable, executable, and not obviously over-constrained or under-specified.
160
161## Step 6: Score sections and deliver the verdict
162
163Use `references/review-rubric.md` to:
164
165- score each reviewed section as `PASS`, `WARN`, or `FAIL`
166- classify findings as `BLOCK`, `CONCERN`, or `OBSERVATION`
167- determine `APPROVE`, `REVISE`, or `REJECT`
168
169Then render a concise founder-review report that includes:
170
171- plan title and mode
172- verdict
173- section summary table
174- blocking issues
175- concerns
176- observations
177- recommended plan changes when revision is required
178
179Do not ask the user procedural questions at the end of the review. Deliver the audit and let the main thread decide what happens next.
180
181**Success criteria**: The verdict follows directly from evidence-backed findings and the rubric.
182
183## Guardrails
184
185- Do not modify the plan file as part of this skill.
186- Do not rubber-stamp a non-trivial plan.
187- Do not invent findings; a clean section can legitimately pass.
188- Do not add `disable-model-invocation`; this skill should remain callable when the user asks for plan review.
189- Do not add `paths:`; this is a generic review workflow.
190- Do not keep giant examples, quality scorecards, or gate tables inline in `SKILL.md`.
191- Do not use `AskUserQuestion` as a substitute for a verdict.
192- Do not forget that this skill runs best as a forked review workflow with separate reasoning budget.
193
194## When To Load References
195
196- `references/personality.md`
197 Use at session start for review persona, traits, and communication style.
198
199- `references/review-checklist.md`
200 Use for section-by-section audit checks and evidence requirements.
201
202- `references/review-rubric.md`
203 Use for mode selection, verdict rules, section scoring, and finding classification (BLOCK/CONCERN/OBSERVATION gate tables).
204
205## Output Contract
206
207Report:
208
2091. reviewed plan and selected mode
2102. verdict
2113. section status summary
2124. blocking issues
2135. concerns
2146. observations
2157. required plan changes when applicable