name: press-release-analyzer
description: Reverse-engineer the PR strategy of a public company from a structured press-release archive (the output of press-release-archiver). Produces a strategic playbook with credibility-ladder analysis, drumbeat patterns, stakeholder orchestration, and foundation-phase plays — designed for early-stage companies (seed → Series A → pre-IPO) who want to learn from how mature companies built their narrative. Supports single-company analysis OR comparative analysis with stage alignment (e.g., "what was Globex Corp doing in its first 5 years post-IPO, vs. Acme Medtech in its first 5?"). Token-conscious: deterministic Python computes stats, then YOU (Claude, in this conversation) synthesize the strategic narrative directly.
Press Release Analyzer
Three deterministic Python scripts produce structured inputs (stats.json, ladder.json, sample.md, etc.). YOU then read those artifacts and write the strategic synthesis directly as markdown. No external API calls — all the LLM work happens in the conversation, charged against the user's Claude subscription.
When invoked
Ask the user for, or infer:
- Archive directory — the output of
press-release-archiver, e.g. acme-medtech/
- Mode — single-company OR comparative (the user will say "compare X and Y" or just give one company)
- CEO/founder names — improves stakeholder-quote attribution. Optional but worth asking about.
- Optional filter — for multi-product-line companies (Abbott, Stryker, Boston Scientific, Johnson & Johnson, etc.) where the archive mixes unrelated business units, ask whether the user wants to narrow the analysis to a specific product line or therapeutic area. If yes, get 3–8 keywords (e.g.
"DBS,deep brain stimulation,Infinity DBS"). The filter applies word-boundary regex matching across headline + summary + body. Filtered output lands in <archive>/analysis-<slug>/ so the same archive can drive multiple parallel playbooks (analysis-dbs/, analysis-diabetes-care/, etc.).
- For comparative: anchor dates for each company. Default = first release in archive. Override with the IPO/SPAC/Series-A date if you want stage alignment.
Stage 1 — Run the deterministic pipeline
For a single-company analysis:
python3 stats.py <archive-dir> --ceo-names "Name1" "Name2"
python3 ladder.py <archive-dir>
python3 sample.py <archive-dir>
This writes:
<archive>/analysis/stats.json — full structured stats<archive>/analysis/patterns.md — human-readable cadence/topic/headline summary<archive>/analysis/ladder.json + <archive>/analysis/ladder.md — candidate credibility-ladder rungs<archive>/analysis/sample.md — strategic-sample bundle of ~30-80 releases
With an optional product-line filter (multi-line conglomerates):
python3 stats.py <archive-dir> --ceo-names "..." --filter "DBS,deep brain stimulation,Infinity DBS"
python3 ladder.py <archive-dir> --filter "DBS,deep brain stimulation,Infinity DBS"
python3 sample.py <archive-dir> --filter "DBS,deep brain stimulation,Infinity DBS"
The filter must be passed identically to all three scripts. Output lands in <archive>/analysis-<slug>/ (the slug is the slugified first keyword, here dbs). Override with --analysis-name <name> to control the subdirectory name explicitly.
patterns.md and stats.json both record the active filter and the match rate (e.g. "73 of 487 releases matched"). The synthesis output you write later should explicitly note that the analysis is narrowed to a product line — frame insights as "the company's DBS communications strategy" rather than "the company's PR strategy."
For a comparative analysis, also run compare.py:
python3 compare.py <archive-a> <archive-b> \
--anchor-a <yyyy-mm-dd> --anchor-b <yyyy-mm-dd> \
--out <archive-a>-vs-<archive-b>/
With a filter (compares only product-line-relevant releases from each):
python3 compare.py <archive-a> <archive-b> \
--anchor-a <yyyy-mm-dd> --anchor-b <yyyy-mm-dd> \
--filter "DBS,deep brain stimulation" \
--out <archive-a>-vs-<archive-b>-dbs/
This writes:
<out>/comparison.json + <out>/comparison.md — stage-aligned cross-company comparison- (When filtered)
<out>/comparison.json also records the match rate for each company. If either company has zero matches the script exits with a clear error.
Important: pick anchor dates that align by company stage, not calendar year. For each company, choose the date it became a public-storytelling entity (typically the IPO date, SPAC merger close, or first material wire release). Aligning these makes the comparison strategically meaningful — what was each company doing at equivalent maturity, regardless of calendar year.
Stage 2 — Read the artifacts
Read in this order:
patterns.md — get the deterministic stats fixed in your head (cadence, topic mix, etc.)ladder.md — see the candidate rungs in chronological ordersample.md — read the actual release content for ~30-80 strategically-selected releases- (For comparative)
comparison.md — see the side-by-side year-by-year deltas
The user is on a Claude monthly subscription, so the cost is your context window, not API tokens. The whole bundle for one company is typically ~15–25K tokens; for a two-company comparison, ~30–50K. Easily fits.
Stage 3 — Write the strategic synthesis
Produce five files in the analysis directory (or the comparison output dir for comparative mode), each as a focused, standalone deliverable. Then concatenate them into a combined playbook.md. Use the Write tool.
File 1 — credibility-ladder.md
The chronological proof-point sequence. Curate the candidate rungs from ladder.md: keep what's strategically meaningful (capital events, regulatory milestones, clinical demonstrations, flagship partnerships); drop the noise (routine appointments unless transformative, building-opening fluff, conference scheduling). For each kept rung:
- Date + headline
- Why it matters strategically (1 sentence)
- Days since previous rung (already computed)
- What the rung enabled / what it set up
End with a Lessons section answering: how were proof-points sequenced? What pattern in spacing? What rung-types dominated each phase? What's the minimum viable ladder an early-stage company could borrow from this?
File 2 — drumbeat-map.md
Cadence patterns and what they reveal. From patterns.md's clusters, silences, and per-year counts:
- Identify the 3–5 most strategically meaningful clusters (drumbeat campaigns) and what they were building toward
- Identify the most meaningful silences (60+ day gaps) — what was being prepared during the quiet?
- Note overall cadence rhythm — steady (Globex Corp-style) or campaign-driven (Acme Medtech-style)?
- For each cluster, list the headlines + your inference of the strategic intent
End with applicable plays for an early-stage client.
File 3 — validator-sequence.md
Stakeholder orchestration. From the stakeholder data in stats.json and the sample releases:
- When did internal voices give way to clinical voices and institutional voices?
- What was the company's first third-party-validated release? What did that signal?
- How was the CEO quote rate used? Always a CEO quote, or selectively?
- What kinds of validators were brought in at what stages?
End with the validator-recruitment playbook: at what milestone do you bring in surgeons / customers / hospitals / board members / investors? What's the right order?
File 4 — foundation-plays.md
Specifically scoped to Years 1–3 of the archive (Foundation→Build phase). For each year:
- What categories of releases did the company prioritize?
- What were the 5–8 dominant play types? (e.g., "headquarters opening," "key hire," "advisory board reveal," "first customer," "TIME's Best Inventions appearance")
- For each play, give a generic template an early-stage client could adapt
This is the highest-value deliverable for the user's stated client base (seed → pre-IPO).
File 5 — playbook.md (combined)
Stitch all four files together with a brief executive summary at the top:
- 3-bullet TL;DR
- "Top 3 plays this archive teaches"
- Then sections for each of the four analyses above
- Final "Caveats and limitations" section
For comparative mode, also produce a sixth file:
File 6 — comparative-playbook.md (only for --compare)
Read comparison.md and write a head-to-head comparison along these dimensions:
- Cadence rhythm (campaign-driven vs steady-state)
- Credibility-ladder pacing (how fast each company climbed)
- Stakeholder strategy (CEO-centric vs distributed voices)
- Topic mix shift over equivalent stages
- Three to five inflection-point plays that one company did and the other didn't (or did differently)
End with a stage-matched recommendations section: which playbook best fits the user's seed/Series-A/pre-IPO clients, and why.
Style guide for synthesis
- Be specific. Quote actual headlines and dates from the archive. Don't say "they used a lot of partnerships" — say "they signed 4 hospital-system agreements in 18 months (Partner A Sept 2023, Partner B Aug 2024, Partner C May 2025, [next] X)."
- Surface contrasts. When stats reveal a delta (e.g., one company has a much higher CEO quote rate than the other, or one has zero clusters while the other has many), call it out. The deltas are the strategy.
- Make it actionable. End each file's analysis section with at least one "play" a client could literally execute next quarter.
- Don't moralize. Note when crisis communications happened (delistings, departures) and analyze the language they used; don't praise or critique.
- Token discipline. You don't need to load full bodies of every release. The sample bundle was specifically curated to give you what you need.
Reporting back
After writing the files, report to the user:
- Which files were written and their paths
- 2–3 most surprising findings from the analysis (the kind of thing they'd want to highlight to a client)
- Any caveats (small-sample issues, classification noise, missing data)
- For comparative mode: the most strategically actionable contrast between the two companies
Failure modes to watch for
- Too few releases — under ~15 releases, statistical patterns are unreliable; surface this caveat.
- Mature company with mostly earnings releases — Globex Corp's archive is dominated by quarterly reporting; the interesting releases are the rare non-earnings events (FDA updates, share repurchases, settlements). Lean into those.
- Anchor mismatch — for comparative, if the two anchor dates aren't actually equivalent in company stage, surface this and recommend a different anchor.
- CEO names not provided — without them, CEO quote rate is undercounted. Ask the user for known CEO/founder names if they didn't provide.
