AGENT SKILL

strategy

Hold the long arc as an auditable record of bets and beliefs — append-only markdown cards under workspace/strategy/ for theses, assumptions, decisions, and signals, cross-linked by [[slug]] so a thesis resolves to the assumptions holding it up and the signals pressing on it. TRIGGER: a strategic bet stated or revised, an assumption named, a competitor or market signal noticed, a consequential decision made, or a request to review the strategy book. SKIP: operational commitments with an owner and a due date (that is ops) and day-to-day choices with no horizon.

License 许可
Apache-2.0
所属仓库
joshuamschultz/Arc
最近更新
2026-10-22
Stars
★ 3
Forks
0

01 前言 Front Matter

name
strategy
version
2.0.0
description
Hold the long arc as an auditable record of bets and beliefs — append-only markdown cards under workspace/strategy/ for theses, assumptions, decisions, and signals, cross-linked by [[slug]] so a thesis resolves to the assumptions holding it up and the signals pressing on it. TRIGGER: a strategic bet stated or revised, an assumption named, a competitor or market signal noticed, a consequential decision made, or a request to review the strategy book. SKIP: operational commitments with an owner and a due date (that is ops) and day-to-day choices with no horizon.

03 SKILL.md 正文

Files

  • strategy/theses/<slug>.md — one card per strategic bet.
  • strategy/assumptions/<slug>.md — one card per assumption a thesis rests on.
  • strategy/decisions/<slug>.md — one card per consequential decision.
  • strategy/signals/signals.md — the single running signal log.

All paths are relative to the agent's own workspace, written with the ordinary file tools (read, write, ls, grep). There is no strategy tool to call and none is needed — this is the agent's own workspace.

Contract

  1. Every thesis, assumption, decision, and signal lands on its card at strategy/<kind>/<slug>.md.

  2. Cards are appended to, never rewritten: read, add the entry at the end, write the whole file back. The superseded belief is the record's whole value.

  3. A new card starts with exactly this header and nothing else:

    # <Title>
    
    - slug: <slug>
    - type: <kind>
    
  4. Each entry is a UTC timestamp heading followed by - key: value lines, omitting every field with no value.

  5. References between cards are written [[slug]], never a bare name.

  6. Every assumption records what would falsify it. An assumption with no falsified_by is an opinion, and is reported back as such.

  7. A challenge appends to the assumption's own card with status: challenged — it never edits the original entry away.

Knowledge

  • Thesis and assumption are separated for falsifiability. A thesis is what is believed; an assumption is the specific claim that, if false, takes the thesis with it. Merged, nothing is ever checkable and a strategy dies of forgetting rather than of being wrong.
  • Strategy fails less from bad thinking than from forgetting — the assumption nobody revisited, the competitor move noticed and dropped, the bet that quietly stopped being true. Appending is what makes revisiting possible.
  • A challenge is evidence, not a verdict. Record it with its source and leave the verdict under review unless one was actually reached.
  • The slug is the identity. Lowercase, non-alphanumeric runs → -, strip ends.
  • Get the timestamp from the clock: run date -u +'%Y-%m-%d %H:%M' and use exactly what it returns. The prompt carries the date but not the time.
  • Never invent a horizon, a confidence, or a source. An empty field is honest.

Fields by kind

Kind Fields
theses name, claim, rationale, horizon, owner, status (default active)
assumptions name, thesis (link), statement, falsified_by, confidence, status: holding
assumptions (challenge) challenge, source, verdict (default under review), status: challenged
decisions name, choice, tradeoff, reversible, thesis (link), revisit_on
signals signal, actor, thesis (link), implication, source

Free-text context goes on a final - note: line.

Steps

To log a thesis, assumption, or decision:

  1. Run date -u +'%Y-%m-%d %H:%M'.
  2. Derive the slug from the name.
  3. read strategy/<kind>/<slug>.md; if absent, start from the header block.
  4. Append a blank line, the ## <timestamp> heading, and one - key: value line per non-empty field, then - note: if there is context.
  5. write the whole file back.
  6. Reply in one line. For an assumption with no falsified_by, add ⚠ nothing would falsify this — it is an opinion, not an assumption.

To challenge an assumption:

  1. Append a new entry to strategy/assumptions/<assumption-slug>.md carrying challenge, source, verdict, and status: challenged.
  2. grep the theses for that assumption's slug and name every thesis resting on it in the reply — the point of the challenge is what it puts at risk.

To log a signal:

Append to the shared strategy/signals/signals.md, linking the thesis it bears on.

To review the strategy book:

  1. ls strategy/theses/, read each, take the last value of each field.
  2. For each thesis, grep the assumptions for its slug.
  3. Report one block per thesis: its latest status and claim, then its assumptions, marking any whose latest status is challenged.
  4. A thesis whose assumptions are all challenged is called out as a bet whose footing has gone.

Red Flags & Rationalizations

Thought Reality
"The thesis and the assumption are the same claim." Then nothing is falsifiable. Split them: the assumption is what can be checked.
"The old belief is wrong now, I'll replace it." The superseded belief is the record's value. Append the new one.
"This assumption is obviously true." Then falsified_by is easy to write. If it isn't, it's an opinion.
"The challenge settles it." Evidence is not a verdict. Leave under review unless one was reached.
"I know roughly what time it is." You know the date, not the time. Run date -u or the entry lands at 00:00.
"The review should show the first status I find." Cards are append-only; the first is the oldest. Take the last.

Validation

A log is correct when:

  • [ ] The card exists at strategy/<kind>/<slug>.md with the exact header block.
  • [ ] Every prior entry is present, unmodified.
  • [ ] The ## YYYY-MM-DD HH:MM heading came from date -u, not 00:00.
  • [ ] Every assumption has a falsified_by, or was flagged as an opinion.
  • [ ] A challenge appended to the assumption card with status: challenged.
  • [ ] Every cross-reference is [[slug]], not a bare name.

A review is correct when every thesis appears once with its latest status and every challenged assumption beneath it is marked.

Examples

An assumption with a falsifier:


## 2026-08-14 17:20
- name: Mid-market buys on compliance
- thesis: [[federal-first-wedge]]
- statement: Mid-market buyers will pay a premium for FedRAMP-grade controls
- falsified_by: Two consecutive quarters where compliance is absent from won-deal reasons
- confidence: medium
- status: holding

Reply: Logged assumption 'Mid-market buys on compliance' under [[federal-first-wedge]]

An assumption with nothing that would falsify it:

Reply: Logged assumption 'AI adoption keeps growing' ⚠ nothing would falsify this — it is an opinion, not an assumption.

A review:

federal-first-wedge — [active] Win federal first, then descend into mid-market
  - mid-market-buys-on-compliance: holding
  - procurement-cycles-shorten: ⚠ challenged (verdict: under review)

04 同仓库更多 Skills