Analyze one or more projects by identifying the top 5 user personas across the full spectrum from technical to non-technical, surfacing gaps each persona hits, and synthesizing what to build next — including breakthrough moves that elevate the project to a new level. Use this skill whenever the user asks "what should I build next", "what do my users want", "what's missing from this project", "who uses this and what do they need", or wants strategic product direction for a repo or portfolio. Trigger even if the user just says "analyze my repos" or "what should I add to X" — persona-based product thinking is almost always the right lens. Works with GitHub URLs, repo names, local project paths, plain project descriptions, or pasted user conversations (emails, Discord, reviews, support tickets).
Installation
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Analyze one or more projects by identifying the top 5 user personas across the full spectrum from technical to non-technical, surfacing gaps each persona hits, and synthesizing what to build next — including breakthrough moves that elevate the project to a new level. Use this skill whenever the user asks "what should I build next", "what do my users want", "what's missing from this project", "who uses this and what do they need", or wants strategic product direction for a repo or portfolio. Trigger even if the user just says "analyze my repos" or "what should I add to X" — persona-based product thinking is almost always the right lens. Works with GitHub URLs, repo names, local project paths, plain project descriptions, or pasted user conversations (emails, Discord, reviews, support tickets).
whats-next
Strategic product direction through persona-based analysis. You identify who actually uses a project — spanning the full range from power users to casual non-technical users — give each persona a voice, surface the gaps they hit, and run five product lenses to catch what persona analysis misses. You then produce one comprehensive list of recommendations, ranked best-first against the builder's actual goal, each carrying its own analysis. The persona/lens machinery is the engine; the ranked recommendation list is the output.
Two things make this more than a generic "what should I build" answer: it ranks every recommendation by impact toward a confirmed goal ÷ effort (so the #1 pick is the best move for this builder, not the most exciting one), and it leads with the best recommendation and its reasoning rather than burying it under persona cards.
After analysis, you offer to action the top pick.
When this skill applies
"What should I build next for [project]?"
"What do my users want?" / "What's missing?"
"Analyze my repos" / "Give me product direction on Y"
Any pasted user conversations, support emails, reviews, Discord logs
Input gathering
Accept whatever the user gives you: GitHub repo names / URLs / local paths / plain descriptions / pasted user conversations / combinations.
If the user gives you nothing specific, ask: "Which project or projects? Or should I look at your GitHub profile as a whole?"
If the user pastes raw user conversations: treat as primary signal — outranks everything else. See references/external-signals.md → "User signal injection."
Confirmation gate — mandatory before any analysis
Before collecting any signal, confirm two things: what you're analyzing and what the user is optimizing for. The goal is what makes "best recommendation" mean something — without it, "best" silently defaults to "most persona-overlap," which is a proxy, not necessarily best for this builder.
Infer a provisional goal from the input and offer it for correction:
A pre-launch / WIP app → launch conversion
A library with issues piling up → reducing maintenance load
A growing tool with adoption signal → retention or revenue
A profile-wide scan → portfolio coherence
Don't know? Offer the two most likely and let the user pick.
Format:
Analyzing:<repo / skill path / description>Optimizing for:<inferred goal> — revenue / adoption / retention / launch conversion / cutting maintenance load / learning / portfolio coherence. Correct me if it's something else.
Mode: Standard / Skill-as-project / Portfolio
I'll start once you confirm.
Wait for confirmation. If the user corrects the goal, re-rank everything against the corrected goal — it changes what counts as "best." The goal is not decoration; it is the ranking key in Phase 3.
Phase 1 — Signal collection
First: detect project type. This is a mandatory gate — run it before any other command.
Step 1a — Filesystem check (local paths only):
ls <given-path>/SKILL.md 2>/dev/null
If the file exists → this is a skill. Stop. Do not continue to Step 1b or signal collection.
Step 1b — Name/description check:
If the input mentions "a skill I built", "a Claude Code skill", a skill name matching the available skills list, or a path that describes a skill — this is a skill. Stop.
If either check triggers:
Read references/skill-mode.md in full. Do not run gh commands. Do not collect signals. Do not form personas. Follow skill-mode.md's signal collection path entirely.
Required output gate — must appear before Phase 2 begins:
> Mode: Skill-as-project (skill-mode.md loaded)
If this line is absent from your output, you have not followed the detection path. Do not proceed to persona identification without it.
If neither check triggers, proceed with standard signal collection:
For everything else, collect from four sources in parallel:
(The README field isn't reliable via --json readme; fetch it separately if needed: gh repo view <owner/repo> | head -40 or read the raw file.) CHANGELOG bugs are especially valuable — a locale bug means real users in that locale; a 50MB import fix means someone hit it.
Description-only input (no repo, not a skill): Sources A–D assume a GitHub repo. When the input is a plain project description (a private/pre-launch app, an idea), you have no repo signal — don't pretend otherwise. Collect what you can: Source C (competitors — always available) and Source B (category conversations on HN/Reddit). Then lean on the pre-launch persona construction in Phase 2 and mark signal quality honestly. A description-only run is legitimately Low or Medium signal; say so.
Source B — External conversations: What people say outside the repo (HN, Reddit, web). See references/external-signals.md. These are the users who didn't care enough to file an issue — often the most representative.
Source C — Competitor landscape: What alternatives exist and how users compare them. Required before writing elevation moves. See references/competitors.md.
Source D — Memory: Check ~/.claude/skills/whats-next-workspace/snapshots/ for a previous run. If found, read the full snapshot file — it is the baseline for the diff. Note the date in the signal header: "previous run from DATE (N days ago)". See references/memory.md for diff format.
Low-signal mode: When signal quality is Low, do not just label and proceed. Execute the Low Signal Protocol:
Label the quality — Signal quality: Low in the signal header.
Mark personas — tag all inferred personas as [Hypothesis].
Generate a validation block — place this section immediately before the "At a glance" table (the table stays strictly last):
## Low Signal — Gather Before Re-Running
Signal quality is Low. The analysis above is hypothesis-based. To get higher-signal output, collect the following before re-running:
**3 validation questions to ask your users / community:**
1. [Domain-specific — e.g. for dev tools: "What made you try this instead of [most obvious alternative]?"]
2. [Gap-targeting — "What's the first thing you'd remove or change if you had 30 minutes?"]
3. [Adoption — "What almost stopped you from trying this?"]
**Signals to gather:**
- [ ] Any GitHub issues (even 1 is more than none)
- [ ] Any HN / Reddit / Discord mentions of the project name or category
- [ ] Any support emails or DMs from users
- [ ] A published GitHub repo (if local-only today)
Tailor the 3 questions to the project domain. Don't generate generic questions.
Phase 2 — Persona identification
Always span most technical → least technical. The non-technical user is the most overlooked.
Archetypes (adapt — don't force):
Power user / developer
Practitioner — serious use, not a developer
Evaluator — hasn't committed yet
Casual / non-technical user
Team deployer — installs it for others
The non-technical floor is mandatory. One of the 5 personas must be the least technical person who can plausibly benefit. This is the most commonly missed persona — and the one whose gaps most often surface recommendations no other persona would generate (plain-language onboarding, trust signals, UX friction that experts stopped noticing).
To find them: ask "who would use this if a technical friend set it up for them?" or "what's the least technical person mentioned in any competitor review?" For a finance tool: a spouse tracking the family budget in a spreadsheet. For a CLI tool: a researcher who follows instructions but doesn't write code. For a messaging app: a family member who just wants to stop missing important group messages.
If the genuinely non-technical persona's only gap is "can't install it" and that's truly unfixable — state it plainly ("This category has no viable non-technical user without a technical intermediary") and explain why. Do NOT silently replace them with a second technical persona. If you end with 5 technical personas and no explanation, you have failed this check. The non-technical user almost always surfaces at least one gap the technical users have stopped noticing.
Avoid overlap. Each persona must drive at least one different recommendation. Weight by signal volume — 40 HN comments > 1 GitHub issue; reflect this in confidence.
Anti-overlap test: Before finalizing, ask: "If I merged Persona A and Persona B into one card, would any recommendation disappear?" If no — merge them and build a genuinely distinct persona instead. Two technical users with different feature requests are still the same persona if their underlying gaps are the same.
Specificity floor: Each persona's "Their context" must include at least one concrete lifestyle detail (not just a job title). Bad: "a developer who wants AI in their workflow." Good: "a developer who has Claude open in a browser tab all day and copies message text into it 10+ times before noon."
For each persona:
### [Persona Name] — [One-line role]
**Evidence:** [source + what it shows — must cite a specific signal, not "plausible" or "assumed"]
**Their context:** [goal + constraints + one concrete daily-life detail]
**Gap they hit:** [specific blocker costing them today — not a wish list item]
**What they'd demand next:** [concrete feature]
Gap ≠ demand. Gap is what's wrong now. Demand is what they want built. They're often at different levels of specificity.
Evidence requirement: If a persona has no signal support (no issue, no commit message, no HN comment, no release title pointing to this user type), mark it explicitly as [Hypothesis] and lower its weight in the ranking.
Pre-launch persona construction: For WIP / unreleased apps with no user data, the non-technical and evaluator personas cannot be grounded in usage evidence. Instead, ground them in launch audience characteristics: What audience is the Product Hunt / launch channel submission targeting? Who shares this type of app on Twitter/Reddit? What are the top App Store reviews for the nearest comparable? Use these as evidence sources when in-app signal is unavailable.
Most dangerous assumption (required, 1–3 sentences)
Before the Missing voices section, name the single biggest assumption in this entire analysis that, if wrong, would invalidate the top recommendation. This is the meta-level perspective the rest of the output can't provide — it's an honest flag to the builder that the analysis rests on a foundation they should verify.
Example: "This entire analysis assumes the core catch-up summary quality is good enough that first users feel it worked. If the summaries are consistently wrong or miss key decisions, the #1 recommendation (driving traffic via a demo) actively hurts — it creates bad first impressions at scale. Verify with 3–5 test users before posting."
One paragraph, not a list. It should be the thing that would make the builder nod and say "yes, that's actually what I'm most uncertain about."
Missing voices (required, 3–5 lines)
After building the 5 persona cards, write a brief "Missing voices" note: which user types are completely unrepresented in the signal? These are the people whose perspective would change the analysis if they spoke.
Examples:
A finance tool with no accountant or tax-professional signal — they'd care about export formats, compliance language, and filing deadlines that none of the current personas raised.
A developer tool with no junior/learner signal — they'd surface onboarding failures that experienced users have forgotten to notice.
A CLI tool with no Windows-user signal — they might find the whole install story broken.
This is not a persona card — it's a named blind spot. One sentence per missing voice: "Enterprise IT buyers: No signal. Would likely prioritize SSO, audit logs, and procurement compliance — all absent from the current recommendation set." This belongs at the end of the personas section. Its function is to tell the builder: here's what might flip the analysis if you gathered it.
Phase 2.5 — Five product lenses (pre-synthesis)
Before writing the recommendations, run these five questions. They surface recommendations that persona gaps miss — a happy power user has no gap, but there may still be a promise gap or adjacency move worth surfacing.
Work through each fully — do not skip a lens. Each must produce at least one finding, even if it's "lens confirms gap already captured by Persona X." If a lens produces nothing, state why explicitly (it becomes "What we set aside" material).
Depth — What would make existing satisfied users use this more? What keeps them from going deeper?
Breadth — What's blocking people who've heard of this but haven't adopted? What would bring in the next 10x of users? For pre-launch / WIP projects: This lens almost always surfaces distribution blockers (no App Store listing, no landing page, no demo video) that persona analysis misses because personas assume users already found the product.
Promise gap — What does the README/description/tagline promise that the project doesn't fully deliver yet? Read the first sentence of the README literally — does the product actually do that?
Adjacency — What do users do immediately before or after this tool? Could this tool absorb those steps and become stickier?
Category leadership — What's the single thing keeping this from being the obvious, default choice in its category? What does the category leader have that this doesn't?
Lens findings feed into the ranked recommendations alongside persona gaps. All 5 lens labels must appear somewhere in the output — either in a recommendation or in "What we set aside." If a lens label is missing from the output, the coverage check fails.
This working section is internal — do not reproduce it in the output. Only the findings appear, labeled in the recommendations where they land.
Phase 2.6 — Domain coverage sweep (anti-omission)
The five lenses are angles; they can all fire and still miss an entire domain that no persona happened to raise. A finance tool with no Security item, an auth library with no threat-model item, a CLI with no install/distribution item — that's a silent hole, and silent holes are how "comprehensive" lies. The sweep is the difference between "what occurred to me" and "what I verified I didn't skip."
Before ranking, walk this checklist. For each domain, either surface a recommendation (it enters the Phase 3 ranking) or consciously decide it doesn't apply. This is fast — most domains resolve in one line.
Domain
The question to ask
Trust (perception)
Does the user understand and believe where their data goes? Is the handling disclosed?
Security (technical)
Is data-at-rest, credentials, and key handling actually protected? (Distinct from trust — a tool can disclose clearly and still encrypt nothing.)
Data & input
Import/export breadth, format coverage, migration, silent data loss / mis-parse
Reliability
Failure modes, error recovery, what breaks under edge input or scale
Performance
Anything slow enough to make a user quit or distrust the result
Accessibility / reach
Who's locked out — by language, disability, platform, or skill level
Distribution / onboarding
How a brand-new user gets from zero to first value
Monetization / sustainability
If revenue or sustainability is the goal: is there any surface for it today?
Compliance / legal
Regulated domain (finance, health, kids, EU data)? Any obligation unmet?
Output discipline: For each domain:
Gap found → ranked recommendation, tagged (from: X domain).
Category-adjacent, no gap found → one line in "What we set aside": "Security domain: tool uses Fernet encryption + audit log — no gap found." The reader needs to know you checked, not just that nothing appeared.
Genuinely orthogonal → appears nowhere. Don't clutter with N/A for domains that don't apply.
The test: if a reader who knows the category would expect a domain to be addressed, silence is a defect — name it either as a rec or as a cleared set-aside line.
Bias check: do not force a domain that doesn't fit just to fill the table, and do not skip one because no persona raised it. The whole point is to catch the domain no persona raised.
Phase 3 — Rank and synthesize
Turn persona gaps + lens findings into one comprehensive list of recommendations, ranked best-first. The rank — not the layer — is the organizing principle of the output. The user wants the single best move at the top with its reasoning attached, then the next best, all the way down. Don't group by type and make the reader hunt; interleave and rank.
Each recommendation carries a type tag so the reader knows what kind of move it is:
Gap — friction actively costing users today. A current user hits this and loses trust or stops. Test: "Would removing this make a user who already found the product stay?" If it needs a new user to show up first, it's a Build, not a Gap.
Build — a new capability. Driven by genuine overlap across personas AND lenses.
Elevation — a bet that changes what the project is, not just makes it better. Test: would you describe the project differently after this move?
The ranking key
Rank across all three types together, by impact toward the confirmed goal ÷ effort. This is the heart of the upgrade:
A cheap Gap fix that directly serves the goal outranks an expensive Build, even though "gaps" sound less exciting than "features." Don't rank by type.
"Impact" means impact toward this builder's goal — not generic value. A feature 3 personas want but that doesn't move the stated goal ranks below a fix only 1 persona wants that unblocks the goal directly.
Within an impact tie, lower effort wins.
State the ranking logic once at the top of the list: "Ranked by impact toward [goal] ÷ effort."
Each recommendation carries its analysis inline
The analysis travels with the recommendation — that is what "best recommendations with analysis first" means. Don't separate the list from the reasoning. For each:
Why it ranks here — its impact toward the goal specifically. This is the sentence that justifies its position.
Source trace — which persona gap(s) and/or lens finding(s) drive it. Only count a persona if their specific gap implies demand for this specific item. A rating without a basis is unverifiable.
Mechanism — concretely what to build or change.
What would change my mind — the one assumption that, if false, drops this item down or off the list. One line. This lets the builder, who knows things you don't, override intelligently.
Success signal — what observable thing in 2–4 weeks tells you this was the right call? One line. Not a metric to optimize — a concrete thing the builder can actually observe without instrumentation if needed. Examples: "3 unprompted mentions of the demo in HN comments," "credential complaints drop out of TestFlight feedback," "widget screenshot appears in a tweet." This turns the recommendation from advice into a testable hypothesis.
Sequencing — required field, ≤6 words: "after #2" or "none." Hard cap. If you're tempted to explain why it depends on #2, that reasoning belongs in "Why it ranks here," not here. And it must be consistent: if #5 says "after #2," then #2's sequencing cannot be "none" — the dependency is mutual information.
Elevation discipline (applies to Elevation-tagged items)
Include at least 2 Elevation moves from different angles: one that changes who the project serves, one that changes how value is delivered (pull → push, sync → async, individual → ambient).
Elevation moves are ranked inline like everything else — they carry the Elevation tag and take their position by impact ÷ effort, never a separate trailing section.
Elevation items carry all the same fields as every other recommendation, with these three added as separate labeled lines (not folded into the prose of "Why it ranks here"):
**Competitor context:** [what rivals do/don't do that makes this structurally yours — must be a discrete labeled field, not prose]
**What would change my mind:** [the falsifiable bet: "Works only if X." Add "Requires: [what to instrument]" if the project can't measure it today.]
**Success signal:** [early observable indicator]
The three Elevation-specific fields are always labeled and always separate — a grader reading the output must be able to find them without parsing prose. Merging them into "Why it ranks here" is a field-completeness failure.
Don't build yet
Specific backfire mechanism per item — not prioritization. "Building Y would require solving Z first, and Z is a harder problem than Y appears" is a reason. "Not a priority" is not.
Phase 3.4 — Goal comparison view
After completing the ranked recommendation list, produce a compact "how this changes under other goals" perspective. This is the most direct answer to "what should I build next" from multiple angles — the same signal, the same gaps, but re-weighted against a different objective.
How to do it:
Identify 2 plausible alternative goals for this project (e.g. if the user confirmed adoption, the alternatives might be retention and cutting maintenance load).
For each alternative, scan the existing recommendation pool and state:
What moves to #1 (or top-3) that wasn't there before
What drops off or becomes "Don't build" under this goal
One sentence on why the answer changes
Keep it compact — this is a movement view, not a full re-run. Each alternative goal gets 3–5 lines.
Why this matters: The same recommendation can be the best move under one goal and actively harmful under another (as the demo showed: "post the launch" was #1 for adoption and "don't build" for cutting maintenance). Showing the builder two alternative lenses lets them sanity-check their goal and spot recommendations that are robust across goals vs. ones that only hold under their specific objective.
Place this section: Between the personas and the diff section (or before the table on first runs). It is output content, not internal machinery.
Phase 3.5 — Coverage check (quality gate)
Before outputting, verify every box is checked. If any fails, fix it before proceeding:
#1 is genuinely the best move toward the goal — not the most exciting or biggest. Does a cheaper item lower down serve the goal better? Then it's #1. The single most important check.
Ranked by goal-impact ÷ effort — types interleaved, not grouped. The At-a-glance table carries the same ★ as each rec, not a collapsed High/Med/Low.
Every rec carries all seven fields — Elevations included — type · impact · effort header, plus Source, Mechanism, "what would change my mind," "Success signal," and Sequencing. Elevations are not exempt from any of these (the most common silent omission). Elevations also add Competitor context + a falsifiable, measurable bet.
Nothing accounted for by silence — scan every persona card's "Gap they hit" field; each must appear in at least one rec's Source line OR in "What we set aside." A gap visible in a persona card that resolves nowhere is the worst kind of omission. Same for all 5 lenses (per lens-coverage line) and every category-adjacent domain (present as rec or cleared set-aside line, never silent).
No phantom citations — every persona named in a Source line has a card in "Personas behind these." Missing Voices personas are NOT valid Source citations — they're blind spots, not evidence. Every lens named in Source actually appears in the lens-coverage line.
No self-narration — the output never references its own machinery. Banned: naming a reference file/line ("per skill-mode.md line 110"), "per the coverage check," "gate skipped," a lens "firing"/"accounted for"/"has no purchase"/"noted rather than forced," naming the Trust-vs-Security split or domain sweep as a mechanic, calling a persona "dead," and naming the output's own sections as scaffolding ("the validation block," "the At-a-glance table tells you…"). State the finding, never that a rule or section produced it. Persona selection is invisible — never "this replaces X." A reader shouldn't be able to tell the skill has phases.
Ratings actually sort — not more than half the recs in the top ★ tier, AND not more than half at the same effort level. If both axes collapse, the ranking is asserted, not computed. Specific trap: if rec #N+1 has equal ★ but lower effort than rec #N, impact÷effort says #N+1 should rank above #N. If you've inverted this, one clause in "Why it ranks here" must explain why (e.g. "after #N because it depends on #N shipping first"). Unexplained inversion = ranking error.
Sequencing is consistent — build the full dependency graph before writing Sequencing fields. If #3 depends on both #1 and #2, write "after #1 and #2" — not just "after #1." If #5 says "after #2," then #2's Sequencing must say "none; #5 depends on this" — not "none" with no acknowledgment. Read every Sequencing field as a pair with the rec it references.
Type minimums: ≥1 Gap, ≥2 Builds with overlap traces, and ≥2 Elevation moves from different angles (one changes who, one changes how) — ranked inline, never a trailing section.
If Signal quality: Low — the "Low Signal — Gather Before Re-Running" validation block appears immediately before the At-a-glance table. Questions must be domain-specific (not generic). A Low-signal run without this block has silently dropped a mandated section — check before closing.
Competitor claims are tagged — any named competitor, named product feature, or "no competitor does X" assertion that you did not directly verify gets [unverified] inline. This applies everywhere: Elevation competitor-context blocks, mechanism descriptions ("competitor Y doesn't have Z"), and comparison tables. The builder needs to know what's checked vs. inferred before publishing any comparison publicly.
"Don't build yet" has ≥1 item with a backfire mechanism (not a priority judgment).
Brand before behavior: if an Elevation makes a positioning claim ("privacy-native"), a Gap fixes any current behavior that contradicts it — else the claim is a credibility trap.
Non-technical persona present — one of the 5 personas is the least technical person who can plausibly benefit. If all 5 are technical users (developers, power users, evaluators-who-write-code), this check fails. The non-technical persona must have a solvable gap that drives at least one recommendation.
Most dangerous assumption stated — the single biggest assumption the analysis rests on, named honestly before Missing voices.
Goal comparison section present — 2 alternative goals named, each with what moves to #1 and what drops off. The recommendations that are robust across all 3 goals are called out.
Missing voices present — at least 2 named user types with zero signal, each with one sentence on what they'd surface that the current personas don't.
Structure: goal in the header matches the gate; recommendations lead; personas + missing voices + goal comparison before the diff; At-a-glance table is last with nothing after it but the footer.
Phase 4 — Diff + save snapshot + offer to action
If a previous snapshot was loaded in Phase 1 Source D, produce a ## What changed since [DATE] section (see references/memory.md for the exact format). This is the most valuable part of a repeat run — skip it only if there is genuinely no previous snapshot.
Save to ~/.claude/skills/whats-next-workspace/snapshots/<slug>-<YYYY-MM-DD>.md. See references/memory.md.
Offer to action the top pick immediately — don't wait to be asked:
"Want me to take #1 further? I can: write a spec, find the relevant code and estimate real effort, draft a validation question for your users, or run a competitor deep-dive."
Output template
# whats-next: [Project Name]
> Mode: Skill-as-project (skill-mode.md loaded) ← include only when in skill-mode; omit for standard repos
**Optimizing for:** [confirmed goal]
**Signal sources:** [repo data, N HN mentions, competitor scan, previous run DATE / first run]
**Signal quality:** High / Medium / Low
---
## Recommendations — best first
*Ranked by impact toward [goal] ÷ effort.*
### 1. [Name] · [Gap / Build / Elevation] · Impact ★★★★★ · Effort M
**Why it ranks #1 (toward [goal]):** [the sentence that earns the top spot]
**Source:** [Persona X gap] + [lens/domain finding]
**Mechanism:** [concretely what to build/change]
**What would change my mind:** [the one assumption that, if false, drops it]
**Success signal:** [what you'd observe in 2–4 weeks that confirms this was right — concrete, not a metric]
**Sequencing:** [≤6 words — "after #3" or "none". Never restate the rationale.]
### 2. [Name] · [Build] · Impact ★★★★☆ · Effort S
**Why it ranks here (toward [goal]):** ...
**Source:** ...
**Mechanism:** ...
**What would change my mind:** ...
**Success signal:** ...
**Sequencing:** none
### 3. [Name] · [Elevation] · Impact ★★★☆☆ · Effort L
**Why it ranks here (toward [goal]):** ...
**Source:** [persona/lens/domain — REQUIRED, same as any rec]
**Mechanism:** [concretely what to build — REQUIRED, elevations are not exempt]
**Competitor context:** [what rivals do/don't do that makes this structurally yours]
**What would change my mind:** [the falsifiable bet — "works only if X." Add "Requires: [what to instrument]" if unmeasurable today.]
**Success signal:** [the early observable indicator — e.g. "privacy community shares the architecture post within 48 hours of publishing"]
**Sequencing:** [terse]
### … [continue — comprehensive, every recommendation, ranked best-first. Every rec — Gap, Build, AND Elevation — carries all six fields: Source, Mechanism, Sequencing are never dropped.]
---
## Don't build yet
- [Thing] — [specific backfire mechanism]
## What we set aside
- [Gap or lens finding] — [why it didn't make the list]
*Lens coverage: Depth → [#N or set-aside] · Breadth → [#N] · Promise gap → [#N] · Adjacency → [#N] · Category leadership → [#N].* ← one terse line; every lens maps to a rec number or "set-aside." This makes a dropped lens impossible to miss.
---
## Personas behind these
*The analytical basis for the ranking above — who these recommendations serve.*
[5 persona cards: Evidence / Their context / Gap they hit / What they'd demand next]
**Most dangerous assumption:** [1–3 sentences naming the single biggest assumption the whole analysis rests on — what, if wrong, would invalidate the #1 recommendation.]
**Missing voices:** [3–5 lines naming user types with zero signal whose perspective would change the analysis. One sentence each: who they are and what they'd surface that the current 5 personas don't.]
---
## If your goal were different
*Same project, same gaps — how the top recommendations reshuffle under 2 alternative goals.*
**Under [Alternative goal A, e.g. retention]:**
- #1 becomes: [what moves to top and why]
- Drops off: [what falls out or becomes "don't build"]
- Why it changes: [one sentence]
**Under [Alternative goal B, e.g. cutting maintenance load]:**
- #1 becomes: [what moves to top and why]
- Drops off: [what falls out or becomes "don't build"]
- Why it changes: [one sentence]
*The recommendations that appear in the top 3 under ALL three goals are your safest bets regardless of which direction you commit to.*
---
## What changed since [DATE]
*(Repeat runs only — omit on first run.)*
**Top recommendations:** [moved up/down, newly added, dropped since last run]
**Still unaddressed from last run:** [items recommended before but not yet built — the most important line]
**Signal changes:** [new external mentions, issues, competitor moves]
**Personas:** [new / dropped / confidence changes]
---
## At a glance
*The scannable index — same order as the recommendations above. Detail is in the sections above.*
| # | What | Type | Impact | Effort | Sequencing | Description |
|---|------|------|--------|--------|------------|-------------|
| 1 | [short name] | Gap / Build / Elev | ★★★★★ | S / M / L | none | [5–8 word why — the problem it solves] |
| 2 | … | | ★★★★☆ | | after #1 | |
Numbered best-first, identical order to the recommendations. **Impact carries the same ★ rating as the recommendation** — not collapsed High/Med/Low. Sequencing shows dependencies so the table works as a standalone planning view — you can see what to ship first and what's blocked without re-reading the full analysis. Effort = S (days) / M (weeks) / L (months).
---
*Snapshot saved. Run again to see what changed.*
Portfolio mode
Prefer cross-tool personas; identify portfolio gaps and sequencing
Apply the five lenses across the portfolio as a whole, not per project
Elevation moves can span the portfolio: unified config, shared auth, meta-orchestration
References
references/skill-mode.md — read first if input is a Claude Code skill
references/external-signals.md — HN/Reddit/web APIs, user signal injection
