Nobody can defend the strategy deck.

The positioning doc written in March is quietly wrong by September — and no one can say which claims died, or why, because you kept the conclusions and threw away the evidence behind them.

Burgess fixes the root cause. It makes your evidence the asset and your strategy docs a regenerable projection of it. Every claim cited. Every artifact rebuildable. Every judgment call left to a human.

Point it at your PRDs, decks, and call transcripts, then ask "why do we claim this?" — and get a real answer, with a source, a date, and a confidence tag.

git diff strategy/projections/positioning.md
# "Here's what changed in our differentiation claims since March —
#  and the exact evidence that moved."

That diff — checkable, dated, sourced — is the thing you can't get from a Google Doc.

The 60-second version

Burgess is a set of skills for AI coding agents (Claude Code and Codex). Install them, open your agent in a repo, and say "set up burgess".

It interviews you, ingests your documents, and builds an evidence base: every customer pain, capability, verbatim quote, and competitive fact stored as a small, dated, source-linked markdown file in git. Your positioning, messaging, and vision are then generated from that evidence — a citation on every claim, an adversarial red-team pass on every draft, and a decision record (with the rejected alternatives) everywhere a human made a call.

New evidence arrives — five sales calls, a lost deal, a churned customer's exit interview — and Burgess computes exactly which conclusions it invalidates and regenerates them. You approve the judgment calls; it does the mechanics.

The very first thing it produces isn't strategy. It's the gap report: an honest account of what your evidence cannot support yet, and the highest-value questions to answer next.

Who it's for: founders writing their first real positioning; product marketers tired of strategy decks nobody can defend; teams who want "why do we claim this?" to have a checkable answer.

Quick start

npx burgess install        # -> ~/.claude/skills (or: --project, --codex, --dest DIR)

(Or manually: unzip burgess-skills.zip into ~/.claude/skills/ or ~/.agents/skills/.)

Then open Claude Code in your product repo (or a fresh one) and say:

"set up burgess"

• Have documents? "Ingest these docs."
• Nothing written down? "Interview me." For early-stage teams the founder interview is ingestion — it's the highest-density evidence source you have.
• Coming back later? "Where were we?" — Burgess reads its manifest and tells you what exists, what's stale, and what needs a human decision.

Why it works this way

Strategy documents rot because they store conclusions, not evidence. Burgess inverts the asset:

• The evidence base is the product. Atomized, source-linked, confidence-tagged, dated.
• Artifacts are projections. Positioning, messaging, web copy: cheap, regenerable renderings of the evidence through established frameworks (April Dunford's positioning, Collins & Porras vision, message house).
• Everything is plain markdown + YAML in git. Diffable positioning is the killer feature: the PR diff is the strategy review.
• Humans decide; the system drafts and argues. Nothing is ever auto-approved.

The pipeline

| Skill | What it does | | ---------------------- | -------------------------------------------------------------------------------------------------------- | | burgess | Orchestrator: scaffolds the repo, detects state, runs steps in order, enforces checkpoints and staleness | | ingest-evidence | Atomizes documents into typed, cited evidence items; regenerates the gap report | | intake-interview | Gap-driven founder/leadership interview; answers become evidence items | | derive-themes | Problem/strength/risk themes; competitive-alternatives map; differentiator vs table-stakes test | | derive-positioning | Dunford five-component positioning; category and beachhead as proposed decision records | | derive-messaging | Message house, per-ICP benefits from real customer language, pitches, objection handling | | generate-vision | Purpose + values discovered from behavior; envisioned future with pillars and milestones | | build-brand-platform | Voice with do/don't pairs; value pillars with proof; the prohibited-claims list | | generate-assets | Web copy, deck outline, internal explainer — message-house claims only | | red-team | Adversarial pass after every regeneration: citation chains, tense scan, skeptic personas |

Start with the wedge (ingest → interview → themes → positioning → red team). Messaging, vision, brand, and assets come when you're ready — positioning churns weekly; brand churns yearly.

What your strategy repo looks like

strategy/
├── evidence/           # THE FOSSIL BED
│   ├── sources.yaml    #   registry of every source (incl. referenced-but-unavailable)
│   ├── items/          #   one file per claim: ev-0001.md ... typed, cited, confidence-tagged
│   └── intake/         #   interview transcripts (append-only)
├── decisions/          # human judgments with rationale + rejected alternatives (dec-001.md ...)
├── projections/        # REGENERABLE — never hand-edited
├── reviews/            # red-team output, one per run (append-only)
├── scripts/check_stale.py   # staleness + citation checker (stdlib Python, CI-ready)
├── gap-report.md       # always current: what the evidence can't support yet
└── manifest.yaml       # the dependency DAG: hashes, staleness, generation records

The git workflow

• Branch per strategy change. New evidence batch or decision → branch → regenerate affected projections → PR. The PR diff is the strategy review.
• Never hand-edit projections. Fix the evidence or the decision, regenerate. check_stale.py detects hand-edits.
• Approval = merge + status flip. A human flips status: approved in frontmatter and fills decided_by on decisions. Agents never do.
• Commit prefixes keep the log readable as a changelog: evidence:, decision:, regen:, review:.
• CI: run python strategy/scripts/check_stale.py — non-zero exit if anything is stale, hand-edited, or has broken citations.

Citation grammar

Projections cite evidence items and decisions; items cite sources. Two hops, each simple.

• [E: ev-0042] — evidence item (carries source, location, confidence, date)
• [D: dec-003] — human decision record
• [I] — inference by the drafting agent (says from what)
• [A] — assumption awaiting confirmation (must appear in open_questions)

A factual sentence in a projection with none of these tags is a defect the red team flags. Secondhand evidence never solely supports a differentiator or headline claim.

Why "Burgess"?

The Cambrian explosion produced a thousand wild body plans — and we only know any of it happened because one evidence bed preserved everything: the Burgess Shale. Every claim paleontology makes about that era traces back to a specimen in that shale, dated and in place.

Your strategy deck is a reconstruction. Burgess is the fossil bed it's reconstructed from. (Hence the trilobite.)

License

MIT