agentleFS
hgambashgambas/think-like-a-pm/.claude/skills/turning-research-into-strategy/SKILL.md

turning-research-into-strategy

Skill0 starsChanged 3 months agoMITView on GitHub

Use when someone has to work out what to build next and wants the answer to hold up. They may have call transcripts, user interviews, painpoint trackers or customer feedback nobody has turned into direction; or a strategy, deck or roadmap whose numbers nobody has checked, or that contradicts another live document; or simply too much going on at once and no clear next step. Works from a spoken brain dump when nothing is written down yet. Not for sparring on a single strategic question, and not for sequencing work inside a direction already settled.

The file

We split this skill into the rules it gives an agent

7 rules. About 0 are common, 7 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈7

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Before: "MRR $5K (committed).

    line 24
  2. Distinctive:

    After: "The deck's $5,000 monthly revenue is committed, not collected — no customer has paid a price yet.

    line 25
  3. Distinctive:

    Sentence-level clarity: Every surviving sentence must survive the question "how do you know?"

    line 32
  4. Distinctive:

    Do NOT use when:

    line 70
  5. Distinctive:

    First word matches a command: read its reference file and follow it.

    line 98
  6. Distinctive:

    After compaction: if the conversation has been summarized since a command's reference file was last read, read that file again before …

    line 102
  7. Distinctive:

    Before: "Should we widen the entry gate or narrow the promise to match the corpus requirement?" After: "Should someone with hardly any …

    line 110

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Turning Research Into Strategy

Turns a pile of customer research into a product strategy where every number carries its receipts. Evidence is inventoried before it is interpreted, goals are counted rather than asserted, and the strategy maps what a customer has connected to what can honestly be promised them.

The Phase Rule

The method runs in ordered phases — Intake → 1 Debrief → 2 Evidence → 3 Reality check → 4 Strategy — each producing its own standalone document, and no phase may run before the ones that precede it exist. Order governs starting, not finishing: any phase may be re-run when new evidence lands, and doubling back is normal — real runs loop. Each phase's document is a real file on disk, not an answer in chat — later phases check for these files, and invoking this skill is the user's explicit request to create them (which satisfies any standing rule against unrequested documentation files). Commands enforce this: every command checks that the documents it depends on exist. If they are missing, it says which, and offers to run the missing phases in order — always asking the user before starting a large pass, never launching one by surprise.

The Finish Condition

Before the first phase runs, agree with the user — in one sentence — what the finished work must be able to do. This is a finish condition for the document, not a product goal (those are derived later, in strategy). The reference shape: "Done when the people building leave the working session knowing what they are prioritizing and why." A sharper one: "Done when the backlog and the integration choice are derived, not opined."

The finish condition may sharpen mid-work as the real problem comes into focus. Re-state it deliberately when it does — record the change in the decision log with its date — and check each phase's output against the current condition before moving on. Work that doesn't serve the finish condition is scope, not progress.

Writing Standard

Applies to every document this skill produces. Write for a smart reader outside the team who was not in the room: everyday words, complete sentences. Any technical or internal term gets a plain-English explanation the first time it appears. Numbers say what they mean ("3 of 20 calls"), people are introduced ("Jules, the founder"), and every sentence can be understood on its own — no abbreviations, arrow chains, or metaphors doing the work of a sentence. When any part of this work is handed to a sub-agent, copy this standard into its instructions verbatim; a summary of it does not survive the handoff.

Before: "MRR $5K (committed). Rev evidence: 0 priced. Deck targets n/a — learning goal only." After: "The deck's $5,000 monthly revenue is committed, not collected — no customer has paid a price yet. We inherit the deck's learning goal, not its revenue targets."

The Compression Finish

The Writing Standard governs sentences as they are written; this governs the finished draft. Once the strategy and Next moves exist, run three passes over the final documents, in order:

  1. Cut self-justification. Any sentence that defends the document rather than informing the reader goes. The receipts stay; the throat-clearing doesn't.
  2. Sentence-level clarity. Every surviving sentence must survive the question "how do you know?"
  3. Structural coherence. Order and headings a stranger can navigate — easy for humans to digest, and plain enough in structure and defined terms that someone else's AI tools can parse it without having been in the room.

The finished document should be smaller than the draft it came from. If the compression pass grew it, it wasn't a compression pass.

Terms Used Throughout

  • Coverage-to-promise table — the strategy's central table: what a customer has connected, what can honestly be answered from it, and what that is good enough for.
  • Mock — a capability the interface implies but that isn't really wired up: visual only, hardcoded, or running on fake data.
  • Source grade — the confidence label stamped on each source at intake: your own transcript outranks a machine transcript, which outranks notes-only.
  • Pre-grouped — a second stamp, separate from the grade and not a rank within it. The grade says how far the words can be trusted; this says the material arrived already sorted into themes by someone else — typically a workshop board, which comes in whole rather than flattened. The stamp travels with the source, and any count drawn from it discloses that it did.
  • Smell — a third stamp, alongside the grade and pre-grouped: something the user believes that nobody has shown yet. The grade says how far the words can be trusted; this says whether the thing described is evidence at all. Most of a spoken brain dump is smells. A smell is excluded from every count in validate, from the goals in strategy, and from every customer-facing promise — and it stays visible throughout as the best guide to what to check next. "Three customers keep asking for this" said out loud is a smell; it becomes evidence when three named customers can be pointed at.
  • Qualifying rule — the test at the top of Next moves that every item must pass to be on it.
  • Next moves — the short document prioritize writes: the ordered work, one page, each item carrying its own reason. Separate from the strategy, because it is the one people forward.

When to Use

Use when:

  • A research corpus exists (calls, painpoints, feedback) and has not been turned into direction — often described by its own owner as "a mess."
  • The team can't translate customer reality into engineering instructions.
  • A strategy already exists but rests on counts nobody has checked, or two live documents contradict each other (different lead goals, two public prices).
  • Build capacity is about to be committed and no one can say which objective each piece of work serves.
  • A prototype exists and nobody has established what it actually does versus what it appears to do.
  • Someone has too much going on at once — signals arriving from every direction — and cannot say what to do first.

Nothing written down yet? Still start here. The material does not have to arrive as documents. Someone living the problem carries most of it in their head, and the fastest way out is to let them talk (see discovery). What they say is stamped smell and never counted — that line is what keeps the numbers honest.

This does not bend the Phase Rule; it is a shorter path through it, and the user is told which one they are on before anything starts:

What existsWhat runsWhat they get
A brain dump and little elsediscovery, then decision-logThe smells written back to them, ordered by what each is holding up, each with what would settle it and who could. Plus the shortlist of material to go and gather. Not a strategy.
A plan or deck, evidence or notchallengeIts core bet named and tested, and the list of claims that need checking.
A real corpusThe full methodStrategy, then Next moves.

Say which row they are on in the first minute, and say what the short path does not produce. A person who was told they would get a strategy and receives a list of questions has been let down, even when the list is the right answer.

Do NOT use when:

  • Direction is settled and the work is sequencing inside it, or writing a single feature spec.
  • The ask is sparring on one strategic question, not producing the document.

Known limits: proven in business-to-business products with a single operator-buyer and a small pipeline of named customers. Untested for consumer products, large pipelines, or markets with no named customers to follow.

Commands

CommandPhaseWhat it doesReference
discoveryIntakeCollect the raw material: transcripts, auto-notes, backlog, deck, prototype access — each graded at the doorreference/discovery.md
challenge [document]Start hereQuestion an existing plan: name its core bet, check it against the evidence, demand receiptsreference/challenge.md
debrief [call]1Unpack one founder call or working session into six bucketsreference/debrief.md
evidence2Flat inventory of the research: attributed, verbatim, uninterpretedreference/evidence.md
validate [claim]2Write the counting rule, count, publish claimed vs. auditedreference/validate.md
teardown3Walk the product screen by screen: working vs. mockup, by trying itreference/teardown.md
feasibility3Audit each data source: what it yields, who grants it, what it costsreference/feasibility.md
strategy4Master command. The coverage-to-promise map, two goals, build rules — runs missing phases firstreference/strategy.md
prioritize4Next moves — the short page you forward. Ordered work under a qualifying rule, each item carrying its own reasonreference/prioritize.md
decision-logAlways onThe working state: decisions locked, open questions, waiting on a named personreference/decision-log.md
checkupUtilityOne-time scan for conflicting skills or instructions on this machinereference/checkup.md

The decision log is not a phase: every command appends to it as it works (decisions it locked, questions it opened, people it is now waiting on). The decision-log command exists to read, tidy, or update it directly.

Routing Rules

  1. No argument: check which of the method's documents already exist in the project (debrief, evidence inventory, validation, teardown, feasibility audit, strategy, Next moves, decision log). Recommend the 2–3 most useful next commands with a one-line reason each, then show the full table. Never auto-run a command from here; the recommendation is a suggestion the user confirms.
  2. First word matches a command: read its reference file and follow it. Everything after the command name is the target. Reading the reference file is non-optional — it defines the flow.
  3. First word doesn't match but the intent clearly maps to one command (e.g. "make sense of this call" → debrief, "is this strategy right?" → challenge, "is this number real?" → validate, "what should we build next?" → prioritize): load that command's reference and proceed as if invoked. If two commands could fit, ask once which.
  4. Missing inputs: if a command's required input is absent — debrief with no transcript, teardown with no access to the product — prompt the user to drop, paste, or connect it before proceeding. Never fill the gap with a guess. prioritize is not one of these: it generates Next moves from the strategy and the teardown, and audits any existing backlog against the qualifying rule rather than waiting to be handed one.
  5. Master run: strategy (and any command whose prerequisite documents are missing) lists what is missing and offers to run the missing phases in order, confirming with the user before each phase.
  6. After compaction: if the conversation has been summarized since a command's reference file was last read, read that file again before continuing the command — a summary of instructions does not survive compaction.

Asking the User

At every decision point that belongs to the user (confirming a counting rule, choosing between candidate goals, approving a phase run), present the choice with the AskUserQuestion tool: a short plain question, 2–4 concrete options, your recommendation first.

Ask in their words, not the method's. The Writing Standard above governs documents; this governs questions, and questions are where it is most often broken. A person cannot answer a question they have to decode. So: no term this skill invented, no word from the trade, no phrase whose meaning depends on having read any of this. Name the real thing at stake and what changes either way. Read it back before sending: could someone who has never seen this skill answer it? If not, rewrite it. If it cannot be asked plainly, it is not understood well enough to ask — go and understand it rather than dressing it up.

Before: "Should we widen the entry gate or narrow the promise to match the corpus requirement?" After: "Should someone with hardly any research be allowed to use this — yes or no?" If this harness has no such tool, ask the same question in plain text and wait for the answer. If no one can answer — an unattended, scheduled, or background run — do not stall: take the recommended option, record it in the decision log as provisional ("taken without confirmation — review me"), and continue. Provisional decisions are surfaced at the next human contact and are never silently promoted to confirmed. Never resolve a user-owned decision silently.

Sub-Agent Dispatches

Heavy passes (inventorying a large corpus, two-pass source verification) may be delegated to sub-agents. Every dispatch must include, verbatim: the Writing Standard above, the counting or grading rules in force, and the instruction to mark inference as inference and unverified claims as unverified.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.