Grounded procurement research for any real-world need. Turn a problem
("reduce bedroom noise from the avenue", "add air conditioning", "hire an
accountant", "buy a TV mount", "replace the kitchen sink") into 3–5 ranked
alternatives spanning DIY-retail → mid-retail → specialty → contractor →
consultant/turnkey, each with cited price bands (low / typical / high),
confidence scores, locale-aware currency, and a final recommendation
with budget. Output is decision-shaped (recommendation + budget envelope),
not research-shaped (knowledge artifact).
Instrucciones de origen · Vista previa de solo lectura
name
procurer
version
0.1.0
description
Grounded procurement research for any real-world need. Turn a problem
("reduce bedroom noise from the avenue", "add air conditioning", "hire an
accountant", "buy a TV mount", "replace the kitchen sink") into 3–5 ranked
alternatives spanning DIY-retail → mid-retail → specialty → contractor →
consultant/turnkey, each with cited price bands (low / typical / high),
confidence scores, locale-aware currency, and a final recommendation
with budget. Output is decision-shaped (recommendation + budget envelope),
not research-shaped (knowledge artifact).
when_to_use
Invoke whenever the user is making a decision that requires calibrated
cost estimates across alternatives. Common triggers:
- "How much would it cost to <fix / acquire / replace / improve> X?"
- "What are my options for <problem>?"
- "Who could I hire to <task>?"
- "Is it cheaper to DIY or get someone?"
- "Give me a budget for <project>."
- User describes a problem and *implicitly* needs a budget to decide.
Use proactively when the user describes a problem without asking for cost
— surface alternatives + bands so they can decide.
not_for
["Pure knowledge research with no decision attached → use `deep-research`.","Technology / library / API selection → use `technical-research`.","Authoring or refining a deployed Life-runtime tenant rules-package (e.g. `materiales-intel.v1` for a constructora tenant). That is module configuration; see the rules-package skeleton at `freelance/_pending-constructora/rules-package/`."]
The procurer skill turns a real-world need into a grounded cost-calibration report with cited alternatives, price bands, confidence, and a recommendation. It is the agent's reflex when the user is about to decide how to spend money or effort on a problem and needs calibration before committing.
Concretely, the output of one procurer pass is:
A clear restatement of the need (and the underlying problem if the user named only a symptom).
3–5 ranked alternatives, each placed on a provider-archetype tier (DIY-retail → mid-retail → specialty product → contractor → consultant / turnkey).
For each alternative: a cost band (low / typical / high) with the currency normalized to the user's locale, plus cited sources (provider URL + page title + fetched-at timestamp) and a confidence score (0–1).
A recommendation — which alternative(s) to pursue, in what order, with the budget envelope.
The output is decision-shaped: the user should be able to read it and act, not need to do further synthesis.
When to invoke (the reflex)
The procurer reflex fires on any of these signals:
Signal
Example
Explicit cost question
"How much does X cost?" / "Give me a budget for Y."
Options request
"What are my options for fixing the noise?" / "Should I get a mini-split or a window unit?"
Hire-someone question
"Who could install this?" / "Should I get a consultant?"
Problem with no cost asked
User describes a problem, no budget request — surface alternatives + bands anyway so they can decide.
Multiple-vendor comparison
"Compare suppliers for varilla #4."
Build-vs-buy / DIY-vs-pro
"Should I do this myself or hire someone?"
Anti-trigger: if the user is researching a topic with no decision attached (e.g., "explain how acoustic windows work"), use deep-research instead. The line: is there a budget envelope at the end?
The 5-stage procedure
Stage 1 — Decompose the need into ranked alternatives
Before any search, restate the need in plain terms and separate the symptom from the failure mode. A "noisy window" is the symptom; the failure mode is usually seal leakage (~90% of the energy on residential sliders) before it's glass mass deficiency. Cheap fixes target the dominant failure mode; expensive fixes replace the system.
Use one of the canonical decomposition patterns (see references/decomposition-patterns.md):
Service that does the work end-to-end with materials it sources.
T5
Consultant / engineer / turnkey
Advisory or full-service end-to-end management.
Not every alternative spans every tier. Capture which tiers are relevant per alternative.
Stage 3 — Choose the search mode
Based on the user's urgency, budget headroom, and decision stakes, choose fast / standard / deep. See references/mode-tiers.md for full contracts.
Mode
Searches
Providers cited per alt
Latency target
When
fast
1 per alt
1–2 (T1 only)
< 1 min total
"Just give me a rough number."
standard
≥3 per alt
3–5 (cover ≥2 tiers)
~3–5 min
Default for most decisions.
deep
≥6 per alt
5–7 (cover ≥3 tiers)
best-effort
"I'm actually deciding now, need the full picture."
Default to standard unless the user explicitly signals "quick" or "thorough."
Stage 4 — Search with grounding discipline
For each (alternative, provider archetype) pair, run grounded web searches. The discipline (see references/grounding-discipline.md) is non-negotiable:
Citation required. Every price must carry source_url, source_title, fetched_at (ISO-8601 UTC). No bare numbers.
No fabrication. If no public price exists for that pair, say so explicitly and leave it empty. Never fill from training data.
Confidence (0–1).
≥ 0.90 — exact product page, SKU + unit explicit.
0.70 – 0.89 — listing exists but unit/spec requires interpretation, or quote-on-request signals.
< 0.70 — category/inferred from comparable item; flag explicitly.
Unit & currency normalization. Quote in the user's locale currency with the right thousand/decimal conventions. Be explicit about tax (IVA/VAT inclusive vs. add).
Sanity bands. If a price falls > 2× the median for that category, flag it in notes — don't drop it (the human decides).
Diversity bias. For standard and deep, prefer at least two tiers — a Tier-1 retail anchor plus a Tier-3+ specialty or contractor benchmark.
Locale-aware suppliers. Use locale-appropriate domains/brands. Default to user's stated region; ask if unclear.
When the agent has a WebSearch / WebFetch tool, use it. When it doesn't, mark the report as unsourced calibration — provide ranges from prior knowledge but explicitly note the absence of fresh citations and recommend the user run a sourced pass before committing.
Stage 5 — Render the report
Produce the report using references/report-template.md. The skeleton:
# <Need restated in one line>
## Problem framing
<2-4 sentences: what's the actual failure mode, not just the symptom>
## Alternatives
### Alternative A — <name> (Tier T1 → T3)
**Thesis**: <what this actually solves>
**Cost band (locale)**: low – typical – high
**Confidence**: 0.X
**Providers cited**: <N> — see footnotes [1] [2] [3]
**Notes**: <lead time, hidden costs, deal-breakers, tax handling>
### Alternative B — ...
### Alternative C — ...
## Cross-cutting notes
- Tax / VAT / IVA treatment
- Locale-specific supplier shortlist
- Common hidden costs
- Lead times
## Recommendation
**Start with**: <Alternative X>
**Total budget envelope**: <low – high>
**Rationale**: <2-3 sentences>
**If that doesn't work**: <Alternative Y as fallback>
## Sources
[1] <url> — <title> — fetched <iso8601>
[2] ...
Then optionally call scripts/validate_report.py <report.md> to lint structural completeness (≥3 alternatives, every alternative has cost band + confidence + ≥1 citation, recommendation present, currency consistent).
Grounding rules (binding)
These rules bind every procurer run regardless of mode:
No price without a citation. A number alone is not procurement research — it's a hallucination risk.
No alternative without a thesis. Don't list options; list options with the problem each solves.
No recommendation without a budget envelope. "Go with X" is incomplete; "Go with X, $A–$B inclusive of installation" is decision-shaped.
No locale assumption. If the user hasn't stated their region, ask before searching — supplier networks and tax handling diverge sharply.
Flag the dominant failure mode early. If 80% of the user's problem can be solved by a 5% intervention (Tier-1 fix), say so before they spend Tier-4 money. Honesty about what actually causes the problem is the most valuable output.
Resources
references/
decomposition-patterns.md — four canonical patterns for breaking a need into alternatives.
provider-taxonomy.md — the 5-tier provider archetype model, with examples across domains.
mode-tiers.md — fast / standard / deep search contracts with budgets and SLA targets.
report-template.md — the report skeleton and a fully-filled exemplar.
scripts/
validate_report.py — structural linter for a generated procurement report. Checks alternatives count, citation completeness, confidence range, recommendation presence, currency normalization. Exit non-zero on failure.
assets/examples/
window-noise-attenuation.md — full worked example: bedroom window noise on a Bogotá avenue, three tiers of remediation.
construction-materials-co.md — generalized Colombian construction materials reference (CO suppliers, families, IVA handling) — lifted from the materiales-intel.v1 rules-package pattern.
Compounding with other skills
deep-research — when the user wants to learn about a topic before deciding, run deep-research first, then procurer for the cost layer.
technical-research — for software/library choice with a cost dimension, do technical-research for the technical evaluation, then procurer for SaaS pricing / consultancy / implementation costs.
bookkeeping (P8) — procurement reports that produce reusable knowledge (e.g., "the CO porcelanato market spans $35k–$120k/m²") should be filed into the entity graph via bookkeeping file.
Closing handoff
The skill ends when the report renders. The agent's last message to the user is the report itself plus a one-line action prompt: "Want me to deep-dive any alternative, refresh citations, or proceed with a specific provider?"
Procurer never spends the money. It calibrates the spend.
